WO2010084839A1 - Likelihood estimation device, content delivery system, likelihood estimation method, and likelihood estimation program - Google Patents

Likelihood estimation device, content delivery system, likelihood estimation method, and likelihood estimation program Download PDF

Info

Publication number
WO2010084839A1
WO2010084839A1 PCT/JP2010/050493 JP2010050493W WO2010084839A1 WO 2010084839 A1 WO2010084839 A1 WO 2010084839A1 JP 2010050493 W JP2010050493 W JP 2010050493W WO 2010084839 A1 WO2010084839 A1 WO 2010084839A1
Authority
WO
WIPO (PCT)
Prior art keywords
likelihood
content
user
input
attribute
Prior art date
Application number
PCT/JP2010/050493
Other languages
French (fr)
Japanese (ja)
Inventor
勇気 小阪
道也 門馬
俊亮 広瀬
Original Assignee
日本電気株式会社
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 日本電気株式会社 filed Critical 日本電気株式会社
Priority to JP2010547477A priority Critical patent/JPWO2010084839A1/en
Publication of WO2010084839A1 publication Critical patent/WO2010084839A1/en

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q30/00—Commerce
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/90—Details of database functions independent of the retrieved data types
    • G06F16/903—Querying
    • G06F16/90335—Query processing
    • G06F16/90344—Query processing by using string matching techniques

Definitions

  • the present invention relates to a technique for estimating the likelihood of a data string for a data model represented by a hidden Markov model, and more particularly to a data sequence expressed by a combination of a symbol and a transition time interval in consideration of the transition time interval. It relates to estimation technology.
  • the data string (or sequence data) is, for example, when characters are input using a keyboard, the input characters are data, and the input characters are arranged in a data string.
  • ⁇ There are two main methods for handling data strings.
  • One is a counting method.
  • the similarity between the data strings is calculated by counting how many pieces of the same data are included and in what order. Furthermore, partial sequences (such as specific character strings) that frequently appear in the entire data string are counted and extracted at high speed.
  • the other is a method of handling a data string using a statistical model. For example, this is a method of assuming that the data string follows a certain probability distribution. Therefore, the distance between the data strings can be measured by the probability distribution probability.
  • the likelihood represents a degree that is likely to occur from the target probability distribution.
  • the data string with time information includes a feature of time interval until data and the next data are observed, this time interval can be used for analysis.
  • the time interval from the data in the data string until the next data is observed is called the transition time interval. If the data string is analyzed in consideration of the transition time interval of the data string, the characteristic may appear in the transition time interval even if no characteristic appears in the data.
  • each user will read content related to his / her interests over a long period of time, while content related to uninterested items will be skipped in a short time. And Therefore, what is the content of each user from the data string that includes the content visit history that the user has visited in what order in the past and the transition time interval indicating the time spent in each content. It can be expected to find more clearly whether you are interested in.
  • the content can be delivered in combination with additional content (such as advertisements) that is tailored to each user's interest. If more effective additional content can be distributed in this way, it can be expected that benefits such as an improvement in advertising revenue will be obtained.
  • additional content such as advertisements
  • the operation history information related to the user's operation is a data string
  • a feature that does not appear only by the data itself of this data string may appear in the transition time interval between the data. It is meaningful to analyze the transition time interval.
  • Non-Patent Document 1 a method for analyzing a data string using a statistical model described in Non-Patent Document 1 and Non-Patent Document 2 will be described.
  • One of the most applied statistical models for handling data strings is a Markov model or a hidden Markov model.
  • the hidden Markov model (Hidden-Markov Model, hereinafter referred to as HMM) is a data model that expresses a data string as a hidden state transition having a data generation model. It is called a “hidden” state in the sense that it is a data string that cannot actually be observed. This model assumes that a data string is generated when a hidden state that is not actually observed transitions.
  • the HMM expresses data as symbols (data values), and is used in fields such as character recognition, natural language processing, speech recognition, and motion recognition.
  • the state where the subscript 1 is attached to the character X is X1
  • the state where the subscript 2 is attached to the character X is X (2)
  • the subscript is attached to the character X.
  • a state in which 1 and superscript 2 are added at the same time is expressed in a form such as X1 (2).
  • Equation 1 A statistical model for modeling a data string by the HMM described in Non-Patent Document 1 and Non-Patent Document 2 is shown in Equation 1.
  • the input data string is X
  • the length of the data X is N.
  • K is the number of HMMs mixed in the data string X.
  • K 1.
  • X ⁇ X1, X2,..., Xn ⁇ .
  • ⁇ k) represents the HMM of the component K.
  • ⁇ k ⁇ k, ⁇ k, Ak, Bk ⁇ is a parameter related to the component K.
  • ⁇ k represents the parameter of the initial probability of the component kth hidden variable.
  • Ak represents a parameter of the transition probability of the component kth hidden variable.
  • Bk represents an occurrence probability parameter of the k-th component symbol.
  • Patent Document 1 describes a technique of performing a search by determining the similarity of moving images from the similarity of HMM.
  • Patent Document 2 describes a technique of modeling customer purchase behavior using an HMM.
  • Patent Document 3 describes a technique of detecting whether a person has fallen by HMM using human image data as input data.
  • Patent Document 4 describes a technique relating to maximum likelihood estimation in which parameters are estimated by an EM algorithm from incomplete data in which a part of data is missing.
  • Patent Document 5 when sensor information that can be acquired from a position sensor, a contact sensor, an audio sensor, a compass sensor, or the like equipped with a wireless device is input, the situation is recognized by the HMM and a web service corresponding to the situation is provided. The technology to do is described.
  • Non-Patent Document 1 and Non-Patent Document 2 describe a technique that considers transition time intervals as an improvement of analysis using only data in a method of analyzing a data string by counting.
  • a transition time interval cannot be incorporated into the model separately from data (symbols).
  • Non-Patent Document 1 and Non-Patent Document 2 ⁇ n is considered multidimensional, and the transition time interval from the (n ⁇ 1) th symbol to the nth symbol is included in the ⁇ n element, The transition time interval can be incorporated into Equation 1 above. However, the transition time interval cannot be modeled separately from the symbol. Furthermore, since this method increases the number of combinations of symbol and transition time interval pairs, calculation time and data capacity increase, which is not practical.
  • Patent Document 1 and Patent Document 2 calculate likelihood using an HMM, but deal with a data series expressed by a combination of a symbol and a transition time interval in question above. It is not what you have.
  • the techniques described in Patent Document 3 and Patent Document 5 analyze the data string by HMM, but do not use the transition time interval.
  • Patent Document 4 only describes a parameter estimation technique using an EM algorithm.
  • An object of the present invention is to provide a likelihood estimation device, a content distribution system, and a likelihood estimation method that enable likelihood estimation as means for effectively discriminating a user attribute based on operation history information related to a user operation. And providing a likelihood estimation program.
  • a likelihood estimation apparatus includes an input unit that receives an input of a data sequence expressed by a combination of a symbol and a transition time interval, and a parameter that represents a specific statistical model of the data sequence.
  • a storage means for storing, a likelihood calculation means for calculating a likelihood for a statistical model of a data series using parameters, and an output means for outputting the likelihood calculated by the likelihood calculation means, To do.
  • a content distribution system is a content distribution system in which a user terminal operated by a user, a web server, and a likelihood calculation device are connected to each other via a network.
  • a web server receives a request from a user terminal, a content receiving unit that distributes content corresponding to the request to the user terminal, a content visit history that indicates a history of the content requested by the user, and a user
  • a user request storage unit that stores a content visit interval from one content to a visit to the next content, and an output that is input to the likelihood calculation device as a data sequence in which the content visit history is a symbol and the content visit interval is a transition time interval
  • a likelihood calculating device Input means for receiving an input of a data series expressed by a combination of a symbol and a transition time interval, a storage means for storing a parameter representing a specific statistical model of the data series, and a likelihood for the statistical model of the data series using the parameter.
  • Input means for receiving a user attribute input calculated by the likelihood calculation device, comprising: a likelihood calculation means for calculating the degree; and an output means for outputting the likelihood calculated by the likelihood calculation means.
  • an advertisement selection means for selecting an advertisement to be added to the content based on the user attribute, and the content distribution means has a function of adding an advertisement to the content and delivering it to the user terminal.
  • another content distribution system is a content distribution system in which a user terminal operated by a user, a web server, and a likelihood calculation device are connected to each other via a network.
  • the web server receives a request from the user terminal, a request accepting unit, a content distributing unit that distributes the content corresponding to the request to the user terminal, a content visit history indicating a history of the content requested by the user, and a user
  • a user request storage unit that stores a content visit interval from one content to a visit to the next content, and a data sequence having a content visit history as a symbol and a content visit interval as a transition time interval are input to the likelihood calculation device.
  • a likelihood calculation device An input means for receiving an input of a data series expressed by a combination of a symbol and a transition time interval, a storage means for storing a parameter representing a specific statistical model of the data series, and a statistical model of the data series using the parameter A likelihood calculating means for calculating likelihood, and an output means for outputting the likelihood calculated by the likelihood calculating means, wherein the web server inputs the attribute of the user calculated by the likelihood calculating device.
  • a page configuration selection unit that selects a page configuration of content based on a user attribute, and the content distribution unit has a function of distributing content to a user terminal by the page configuration .
  • a likelihood calculation method is a likelihood of calculating a likelihood for a specific statistical model of a data sequence for a data sequence expressed by a combination of a symbol and a transition time interval.
  • a calculation method in which an input unit receives an input of a data series, reads parameters of a statistical model from a storage unit, and a likelihood calculation unit calculates a likelihood for the statistical model of the data series using the parameters, and is calculated The likelihood is output to the output means.
  • a likelihood calculation program calculates a likelihood for a data sequence expressed by a combination of a symbol and a transition time interval, with respect to a specific statistical model of the data sequence.
  • a calculation device a procedure for receiving an input of a data series, a procedure for reading a parameter of a statistical model, a procedure for calculating a likelihood for the statistical model of the data series using the parameter, and a means for outputting the calculated likelihood And outputting the data to the computer.
  • the present invention is configured to calculate the likelihood for a known statistical model in consideration of the transition time interval for the data series expressed by the combination of the symbol and the transition time interval as described above, the user's Based on the operation history information related to the operation, likelihood estimation as means for effectively determining the attribute of the user can be performed.
  • FIG. 3 is an explanatory diagram showing the configuration of the likelihood calculating apparatus according to the second embodiment of the present invention. It is a flowchart shown about the operation
  • FIG. 6 is an explanatory diagram illustrating an example of a content visit history X and a content visit interval T that are input from the input unit to the likelihood calculating unit and the attribute model learning unit illustrated in FIG. 5. It is a flowchart showing the attribute model learning process which the attribute model learning means shown in FIG. 5 performs. It is a flowchart showing the estimation process of the user attribute which the likelihood calculation means shown in FIG. 5 performs. It is explanatory drawing which shows the example of the likelihood with respect to each attribute calculated by step S903 of FIG. It is explanatory drawing which shows the example of the output of a user's attribute output by step S905 of FIG.
  • FIG. 17 is a sequence diagram showing an operation of an attribute model learning process performed by the user terminal, the web server, and the attribute estimation device shown in FIGS.
  • FIG. 17 is a sequence diagram illustrating an example of an operation of an advertisement distribution process corresponding to an estimated attribute performed by the user terminal, the web server, and the attribute estimation device illustrated in FIGS.
  • FIG. 25 is a flowchart illustrating likelihood calculation and spoofing detection processing performed by the likelihood calculation unit and the spoofing detection unit illustrated in FIG. 24. It is explanatory drawing which shows the example of the determination data output via the output means by the process shown to step S2602 of FIG.
  • the likelihood calculation apparatus 100 is a likelihood calculation apparatus that calculates the likelihood of a specific statistical model of a data series for a data series expressed by a combination of a symbol and a transition time interval.
  • An input unit 101 that receives an input of a data sequence, a storage unit (HDD 104) that stores a parameter 111 of the statistical model, a likelihood calculation unit 110 that calculates a likelihood for the statistical model of the data sequence using the parameter 111,
  • Output means 105 for outputting the calculated likelihood.
  • the statistical model is expressed by either a hidden Markov model or a Markov model
  • the transition time interval is expressed by either a discrete value or a continuous value.
  • the present embodiment can calculate the likelihood for the statistical model represented by the parameter 111. Hereinafter, this will be described in more detail.
  • FIG. 1 is an explanatory diagram showing a configuration of a likelihood calculating apparatus 100 according to the first embodiment of the present invention.
  • the likelihood calculation device 100 is a general computer device, and includes an input unit 101, a CPU (Central Processing Unit) 102, a RAM (Random Access Memory) 103, an HDD (Hard Disk Drive) 104, and an output unit 105.
  • the input unit 101 is a device that allows a user to perform data input operations, such as a general keyboard.
  • the CPU 102 is a main body that executes a computer program, and each program described later is executed here.
  • the RAM 103 is a volatile storage device that stores each program executed by the CPU 102 and temporary storage data.
  • the HDD 104 is a non-volatile storage device that accommodates each program and data when not executed.
  • the output means 105 is a device that can show a calculation result to the user, such as a general display.
  • the likelihood calculating means 110 is constructed on software by the CPU 102 reading a computer program from the HDD 104 into the RAM 103.
  • the likelihood calculating means 110 associates the data string with the statistical model and calculates the likelihood of the data string for the statistical model.
  • the likelihood here represents the similarity between the data string and the model.
  • this statistical model is prepared in advance in the HDD 104 as the parameter 111.
  • N the total number of data in the data string input by the user
  • Xi the i-th (1 ⁇ i ⁇ N) data among them
  • Ti the i-th transition time interval
  • the likelihood calculation means 110 calculates the likelihood P (X, ⁇ ) for the model for each of these input data strings, and outputs the calculation result to the output means 105.
  • FIG. 2 is a flowchart showing the operation of the likelihood calculating means 110 shown in FIG.
  • the likelihood calculating means 110 first accepts input of data strings X and T from the input means 101 (step S201). Subsequently, the likelihood calculating unit 110 reads the parameter 111 from the HDD 104 (step S202), and calculates the likelihood of the data strings X and T using the following equations 3 and 4 (step S203). Then, the calculation result is output to the output means 105 (step S204).
  • the input data string is X
  • the length of the data X is N.
  • the likelihood calculation of the data strings X and T performed in step S203 is expressed by the following equation.
  • ⁇ k represents a parameter of the occurrence probability of the component k-th transition time interval.
  • X represents a symbol string of the data string.
  • N be the length of the data string.
  • T represents a sequence of transition time intervals of the data sequence.
  • a likelihood calculation method is a likelihood calculation method for calculating a likelihood for a specific statistical model of a data sequence for a data sequence expressed by a combination of a symbol and a transition time interval
  • the means accepts the input of the data series (FIG. 2: step S201), reads the parameters of the statistical model from the storage means (FIG. 2: step S202), and the likelihood calculating means uses the parameters and the likelihood for the statistical model of the data series Is calculated (FIG. 2: step S203), and the calculated likelihood is output to the output means (FIG. 2: step S204).
  • each of the above-described operation steps may be programmed to be executable by a computer, and may be executed by the likelihood calculating apparatus 100 which is a computer that directly executes each of the steps.
  • the present embodiment has the following effects. According to this embodiment, it is possible to estimate the likelihood considering the transition time interval for the statistical model previously given as the parameter 111.
  • the probability that the transition time interval has occurred from the nth hidden state may be a value obtained by discretizing the time interval.
  • the probability represented by the following formula may be a multinomial distribution.
  • the probability that the nth transition time interval of the mth data string has occurred from the nth hidden state in the kth component shown in Equation 5 is a continuous value as shown in the following equation. It is good. In this case, the probability indicated by the left side of the following equation may be a continuous distribution.
  • the following formula is for an exponential distribution.
  • ⁇ i (k) is a parameter of the component k-th exponential distribution.
  • the meanings of the other symbols are the same as those in the first embodiment.
  • the likelihood calculating means 103 may perform calculation using the following expression instead of the expression 3. This is a model in which the time interval ⁇ n arises from the hidden states Sn and Sn-1.
  • ⁇ n may be a continuous value or a discrete value.
  • the element shown by the following formula may be a multinomial distribution.
  • the time interval may be a polynomial.
  • the probability indicated by the left side of the following equation may be a continuous distribution.
  • the following formula is for an exponential distribution.
  • the data time interval is incorporated into the hidden Markov model.
  • the data time interval may be incorporated into a Markov model other than the hidden state.
  • the following equation shows a model that incorporates the data time interval into the Markov model.
  • P ( ⁇ n) represents the probability that the component k-th ⁇ n will occur.
  • ⁇ n ⁇ 1, Ak) represents the probability that ⁇ n transitions from the component kth ⁇ n ⁇ 1.
  • P ( ⁇ n) (k) represents the probability that the component k-th ⁇ n will occur.
  • ⁇ n ⁇ 1) represents the probability of transition from ⁇ n ⁇ 1 to ⁇ n of the component kth.
  • the likelihood calculating apparatus 300 further estimates and stores the parameters of the statistical model using the data series input to the likelihood calculating apparatus 100 according to the first embodiment.
  • the parameter estimation means 320 to be stored in is added.
  • the present embodiment can learn the parameters of the statistical model from actual data, and can more accurately calculate the likelihood. Hereinafter, this will be described in more detail.
  • FIG. 3 is an explanatory diagram showing a configuration of a likelihood calculating apparatus 300 according to the second embodiment of the present invention.
  • the likelihood calculation device 300 is a computer device similar to the likelihood calculation device 100 according to the first embodiment shown in FIG. 1, and has the same input means 101, CPU 102, RAM 103, HDD 104, and the like as the likelihood calculation device 100.
  • Output means 105 is provided.
  • the functions and operations of the operation units called with the same names and reference numbers are the same as those described as the first embodiment.
  • the likelihood calculating means 110 and the parameter estimating means 320 are constructed on software by the CPU 102 executing a computer program.
  • the parameter estimation unit 320 estimates the parameter 111 from these data strings and stores it in the HDD 104.
  • the likelihood calculating means 110 reads the parameter 111 and calculates the likelihood of the data string for the model by the same operation as described in the first embodiment.
  • FIG. 4 is a flowchart showing an operation in which the parameter estimation unit 320 shown in FIG.
  • the parameter estimation unit 320 receives input of the data strings X and T from the input unit 101 (step S401). Subsequently, the parameter estimation unit 320 models the data X and T using the above-described equation 1, and estimates the parameter 111 that maximizes P shown in the following equation using a known EM algorithm (step S402). The parameter estimation unit 320 writes and stores the parameter 111 estimated by this calculation in the HDD 104 (step S403).
  • the meanings of the other symbols are the same as those in Formula 1 and the first embodiment.
  • the likelihood calculating means 110 reads the parameter 111 and calculates the likelihood of the data string for the model by the same operation as described in the first embodiment. This likelihood calculation may be performed simultaneously with the operation of estimating the parameter 111 shown in FIG. 4 or may be performed at another timing.
  • the parameter estimation means estimates the parameters of the statistical model using the input data series (FIG. 4: steps S401 to 402). Then, an operation of storing the estimated parameters in the storage means by the parameter estimation means (FIG. 4: step S403) is added.
  • each of the above-described operation steps may be programmed to be executable by a computer, and may be executed by the likelihood calculating apparatus 300 which is a computer that directly executes each of the steps.
  • the present embodiment has the following effects. According to this embodiment, it is possible to learn the parameter 111 of the statistical model from actual data, and it is possible to calculate the likelihood more accurately.
  • the probability that the n-th transition time interval of the m-th data string has occurred from the n-th hidden state in the k-th component of the equation 1 used in the above-described step S402 is the discrete time interval symbol. You may obtain
  • Equation 6 (Modification 3 of the second embodiment) Further, the equation shown in Equation 6 may be used instead of Equation 1 in Step S402 described above.
  • the n-th transition time interval ⁇ mn of the m-th data string which is the element of Equation 6, is the n ⁇ 1-th hidden state Sn ⁇ 1 and n of the k-th component.
  • the probability of occurrence from the first hidden state Sn can also be obtained by the equation shown in Equation 7.
  • the third embodiment of the present invention is a likelihood calculation device (attribute estimation device 500) according to the first and second embodiments, and the content visit history in which the input means 101 indicates the history of web content visited by the user. , And a content visit interval from the time when the user visits one web content to the next visit to the next web content as transition time intervals. Then, the likelihood calculating means 510 outputs the calculated likelihood as a user attribute.
  • this symbol can include a history of search words used when the user searches for web contents.
  • the present embodiment can calculate the attribute of each user of web content as likelihood. Hereinafter, this will be described in more detail.
  • the likelihood calculation apparatus is a content (web page) or additional content (web page) suitable for each user by the web server.
  • This is an embodiment in which it is used as an attribute estimation device that estimates the user's attributes (preference, hobby, interest, age, sex, occupation, etc.) in order to deliver advertisements or links to other contents.
  • the system for managing web content selects the content to be distributed as additional content according to the user attributes. For example, an advertisement of a product similar to a product previously purchased by the user is distributed as additional content. In many cases, the operating cost of the system is derived from this advertising cost.
  • the content visit history indicating which content the user has visited in the order and the content visit interval from visiting a certain content until visiting the next content is input.
  • FIG. 5 is an explanatory diagram showing a configuration of an attribute estimation apparatus 500 according to the third embodiment of the present invention.
  • the attribute estimation apparatus 500 is a computer apparatus similar to the likelihood calculation apparatus 300 according to the second embodiment shown in FIG. 3, and has the same input means 101, CPU 102, RAM 103, HDD 104, output as the likelihood calculation apparatus 300. Means 105 is provided.
  • the functions and operations of the operation units called with the same names and reference numbers are the same as those described as the first and second embodiments.
  • the likelihood calculation means 510 and the attribute model learning means 520 are constructed on software by the CPU 102 executing a computer program. As described above, a data string composed of M symbols and M transition time intervals is input from the input unit 101 to these operation units as in the first and second embodiments.
  • Likelihood calculation means 510 and attribute model learning means 520 operate in the same manner as likelihood calculation means 110 and parameter estimation means 320 in likelihood calculation apparatus 300 according to the second embodiment. Then, the learned model parameter 111 is written in the HDD 104. For each attribute model, Equation 3 described in the first embodiment is used. And the parameter 111 for every attribute is (theta) and (theta) (K) demonstrated in step 402 of FIG. 4 of 2nd Embodiment.
  • FIG. 6 is an explanatory diagram showing an example of model parameters learned for each attribute by the attribute model learning means 520 shown in FIG.
  • the process of learning the parameter ⁇ representing the model for each attribute that is a candidate for estimation is referred to as an attribute model learning process.
  • the likelihood calculating means 510 reads the parameter 111 necessary for estimating the attribute from the HDD 104 by the attribute that is the estimation candidate, and extracts the user's attribute. presume. Then, the estimated user attribute is determined from the candidates, and the result is output to the output means 105.
  • FIG. 7 is an explanatory diagram showing an example of the content visit history X and the content visit interval T input from the input unit 101 to the likelihood calculating unit 510 and the attribute model learning unit 520 shown in FIG.
  • the content visit interval T is treated as a discrete symbol such as MID or LOW. For example, when the visit interval is within 1 minute, LOW is defined, and when the visit interval is within 1 hour, MID is defined in advance.
  • the arrow ( ⁇ ) indicating the content visit history in FIG. 7 represents a transition from one content to the next content. Furthermore, an arrow ( ⁇ ) indicating a content visit interval represents a transition of a visit interval from a certain content until a next content is visited.
  • the parameter 111 stored in the HDD 104 is the parameter ⁇ of each attribute model learned by the attribute model learning means 520.
  • the user attributes estimated by the likelihood calculation means 510 are finally output to the output means 105.
  • FIG. 8 is a flowchart showing the attribute model learning process performed by the attribute model learning means 520 shown in FIG.
  • the attribute model learning unit 520 receives the input of the data strings X and T (step S801).
  • the attribute model learning unit 520 reads the attribute model parameter 111 corresponding to the input user attribute from the HDD 104 (step S802).
  • the initial value parameter is read out.
  • the attribute model learning unit 520 learns the attribute model using the input content visit history X of the user, the content time interval T, and the parameter 111 (step S803).
  • the learning of the attribute model here is synonymous with the parameter estimation described in step S402 in FIG.
  • the equation 3 described in the first embodiment is used as the attribute model of each attribute model, and the parameter 111 that maximizes P shown in the equation 9 is estimated using a known EM algorithm.
  • the attribute model learning unit 520 writes and stores the parameter 111 estimated by this calculation in the HDD 104 (step S804). Next, attribute estimation processing will be described.
  • FIG. 9 is a flowchart showing a user attribute estimation process performed by the likelihood calculating means 510 shown in FIG.
  • the likelihood calculating means 510 receives the same data strings X and T as those inputted to the attribute model learning means 520 in FIG. 8 (step S901). For this data string, the likelihood calculation means 510 reads the parameter 111 of the attribute model that is a candidate for attribute estimation (step S902), and calculates the likelihood P (X, T) for each attribute model (step S903). .
  • This likelihood calculation is the same as the content of the calculation performed by the likelihood calculating means 103 of the first embodiment in step S203 of FIG.
  • FIG. 10 is an explanatory diagram showing an example of likelihood for each attribute calculated in step S903 in FIG. For example, if genre names representing user preferences such as “sports”, “animation”, “news”, “travel”, and the like are attribute candidates, the estimated content visit history and the content visit interval of the user are input. The likelihood is output as shown in FIG.
  • the likelihood calculating means 510 estimates the user attribute from the likelihood calculated in step S903 (step S904).
  • the attribute of the attribute model that outputs the highest likelihood can be used as the estimation result of the user.
  • the estimation result is “sport”.
  • the likelihood calculating unit 510 outputs the estimated user attribute to the output unit 105 (step S905).
  • FIG. 11 is an explanatory diagram showing an example of the output of the user attribute output in step S905 of FIG. In the example illustrated in FIG. 11, an output indicating that the user attribute is “sports” is output. (Overall operation of the third embodiment) Next, the overall operation of the above embodiment will be described.
  • the operation according to the present embodiment is the operation of the first and second embodiments, and the likelihood calculating means 110 outputs the calculated likelihood as a user attribute.
  • the above-described operation steps may be programmed to be executable by a computer, and may be executed by the attribute estimation apparatus 500 which is a computer that directly executes the respective steps.
  • the present embodiment has the following effects. According to the present embodiment, attributes such as user preferences, hobbies, interests, age, sex, occupation, etc. can be accurately estimated and used for content distribution.
  • the content visit time interval T input from the input unit 101 may be a continuous value instead of a discrete symbol.
  • FIG. 12 is an explanatory diagram illustrating an example of the content visit history X and the content visit interval T input from the input unit 101 in the present modification.
  • the content visit time interval T is not a discrete symbol such as MID or LOW but a specific number of seconds.
  • the search word history can also be used for user attribute estimation.
  • the search word here is a word that becomes a search keyword input to a search engine such as Yahoo (registered trademark) or Google (registered trademark) when the user searches the content.
  • FIG. 13 is an explanatory diagram illustrating examples of the content visit history X, the content visit interval T, and the search word that are input from the input unit 101 in the present modification.
  • the attribute estimation device 500 may be configured such that the attribute model learning unit 520 is omitted.
  • the fourth embodiment of the present invention is a content distribution system that includes the attribute estimation apparatus 500 according to the third embodiment described above, and the web server determines the user's attribute according to the user attribute estimated there.
  • advertisement information corresponding to an attribute is added to the content and distributed.
  • a content distribution system 1400 includes a user terminal 1401 operated by a user, a web server 1403, and an attribute estimation apparatus 500 according to the third embodiment. This is a content distribution system connected to each other via (Internet 1402).
  • the web server 1403 includes a request reception unit 1612 that receives a request from the user terminal 1401, a content distribution unit 1614 that distributes content corresponding to the request to the user terminal, a content visit history that indicates a history of the content requested by the user, and A user request storage unit 1623 for storing a content visit interval until a user visits the next content from one content, and a likelihood calculation device as a data series having a content visit history as a symbol and a content visit interval as a transition time interval 500, an input unit 1601 that receives an input of a user attribute calculated by the likelihood calculating device, and an advertisement selection unit that selects an advertisement to be added to the content based on the user attribute 161 Has a door, content delivery means 1614 is delivered to the user terminal by adding advertising to the content.
  • the present embodiment can distribute advertisements with contents suitable for the calculated attributes of each user to the user terminals. Hereinafter, this will be described in more detail.
  • the attribute estimation apparatus 500 which concerns on 3rd Embodiment is utilized for estimation of a user's attribute.
  • FIG. 14 is an explanatory diagram showing a connection configuration of each device in the content distribution system 1400 according to the fourth embodiment of the present invention.
  • This content distribution system includes a user terminal 1401 which is a computer device that provides an input / output interface to a user, and content obtained by adding advertisement information corresponding to a user attribute to the content requested by the user.
  • a web server 1403 that is a computer device distributed to a user and an attribute estimation device 500 according to the third embodiment are configured to be connected to each other via the Internet 1402.
  • FIG. 15 is an explanatory diagram showing the configuration of the user terminal 1401 shown in FIG.
  • the user terminal 1401 is a computer device such as a general personal computer, and includes an input unit 1501, a CPU 1502, a RAM 1503, a network card 1504, and an output unit 1505.
  • the network card 1504 is an interface that communicates by connecting to the Internet 1402.
  • the functions and operations of the respective components other than these are the same as the functional units having the same names in the likelihood calculation device and the attribute estimation device described with reference to FIGS. 1, 3, and 5.
  • the content request unit 1511 and the content display unit 1512 are constructed on software by the CPU 1502 executing a computer program.
  • the content request unit 1511 receives an input about the content that the user wants to visit through the input unit 1501, and requests this request to the web server 1403 through the network card 1504. For example, when a user wants to visit sports content, he / she requests the sports content from the web server.
  • the content display unit 1512 receives the content returned from the web server 1403 in response to the request via the network card 1504 and displays it on the output unit 1505.
  • FIG. 16 is an explanatory diagram showing the configuration of the web server 1403 shown in FIG.
  • the web server 1403 is a computer device such as a general server computer, and includes an input unit 1601, a CPU 1602, a RAM 1603, an HDD 1604, a network card 1605, and an output unit 1606.
  • the web server 1403 executes a function of adding advertisement information corresponding to the user attribute to the content requested from the user terminal 1401 and delivering the advertisement information to the user terminal 1401.
  • the network card 1504 is an interface that communicates by connecting to the Internet 1402, similar to the user terminal 1401 of FIG. 15.
  • the input unit 1601 and the output unit 1606 input / output information to / from the attribute estimation apparatus 500.
  • any connection method can be applied.
  • the functions and operations of the respective components other than these are the same as the functional units having the same names in the likelihood calculation device and the attribute estimation device described with reference to FIGS. 1, 3, and 5.
  • the advertisement selection unit 1611, the request reception unit 1612, the reading unit 1613, and the content distribution unit 1614 are constructed on software by the CPU 1602 executing a computer program.
  • the HDD 1604 includes storage units for information such as an advertisement storage unit 1621, a content storage unit 1622, and a user request storage unit 1623. Hereinafter, each content will be described.
  • the advertisement selection unit 1611 receives the user attribute estimation result from the user attribute estimation apparatus 500 and extracts the advertisement corresponding to the user attribute corresponding to the estimation result from the advertisement storage unit 1621. Then, the advertisement extracted from the advertisement storage unit 1621 is added to the content distributed from the content distribution unit 1614 to the user.
  • the request accepting unit 1612 accepts a content request from the user terminal 1401.
  • the reading unit 1613 reads from the user request storage unit 1623 the past content visit history and content visit interval of the user who made the request.
  • the content distribution unit 1614 distributes the content requested from the user terminal 1401 to which the advertisement information corresponding to the user attribute is added to the user terminal 1401.
  • the user request storage unit 1623 stores the content requested by the user together with the user name and the requested time.
  • the content storage unit 1622 stores the content distributed to the user by the content distribution unit.
  • the advertisement storage unit 1621 stores advertisements distributed by the advertisement selection unit 1611.
  • the input unit 1601 passes the user attribute estimation result input from the attribute estimation device 500 to the advertisement selection unit 1611.
  • the output unit 1606 outputs the user's past content visit history and content visit interval extracted by the reading unit 1613 to the attribute estimation apparatus 500.
  • the attribute estimation apparatus 500 has the same configuration as that shown in FIG. 5 and performs the operation already described as the third embodiment.
  • FIG. 17 is a sequence diagram showing the operation of the attribute model learning process performed by the user terminal 1401, the web server 1403, and the attribute estimation apparatus 500 shown in FIGS.
  • a user whose attribute is known in advance operates the content request unit 1511 of the user terminal 1401, and requests content from the web server 1403 (step S1701).
  • the request reception unit 1612 receives this request and stores it in the user request storage unit 1623 (step S1702).
  • the content distribution unit 1614 reads the requested content from the content storage unit 1622, and the content distribution unit 1614 distributes it to the user terminal 1401 (step S1703).
  • the content display unit 1512 receives and displays the distributed content (step S1704).
  • the content distribution unit 1614 outputs the information of the user who accepted the request to the reading unit 1613 (step S1705).
  • the “information of the user who accepted the request” here is information that can read the user's past content visit history and content visit interval from the user request storage unit 1623, and the user and the history can correspond one-to-one. I just need it.
  • the reading unit 1613 reads from the user request storage unit 1623 the past content visit history and content visit interval of the user who accepted the request (step S1706). Furthermore, the reading unit 1613 inputs the read content visit history of the user, the content visit interval, and the previously determined user attribute to the attribute estimation apparatus 500 through the output unit 1606 (step S1707).
  • the attribute estimation apparatus 500 reads the parameter 111 of the attribute model corresponding to the input user attribute from the HDD 104. If the attribute model corresponding to the input user attribute has never been learned, the initial value parameter 111 is read from the HDD 104.
  • the attribute model learning means 520 learns an attribute model using the input past content visit history, content time interval, and parameter 111 of the user (steps S1708 to 1709). Then, the parameter 111 based on the learning result is written in the HDD 104 (step S1710). This completes the attribute model learning process.
  • FIG. 18 is a sequence diagram illustrating an example of the operation of the advertisement distribution process corresponding to the estimated attribute performed by the user terminal 1401, the web server 1403, and the attribute estimation apparatus 500 illustrated in FIGS.
  • the user operates the content request unit 1511 of the user terminal 1401 to request content from the web server 1403 (step S1801).
  • the request receiving unit 1612 receives this request and stores it in the user request storage unit 1623 (step S1802).
  • the content distribution means 1614 outputs the information of the user who accepted the request to the reading means 1613 (step S1803).
  • the “information of the user who accepted the request” here is as described in step S1705 of FIG.
  • the reading unit 1613 reads the user's past content visit history and content visit interval from the content request recording unit 1602 (step S1804), and inputs them to the attribute estimation apparatus 500 through the output unit 1606 (step S1805). .
  • the likelihood calculation means 510 reads the parameters 111 of all attribute models (step S1806), and outputs the likelihood for the input content visit history and content visit interval from each of these attribute models (step S1806). S1807). Then, the user's attribute is estimated from the likelihood output from each attribute model, and this is input to the advertisement selection means 1611 through the input means 1601 (step S1808). Among the likelihoods output by each attribute model, the attribute of the attribute model that outputs the highest likelihood is used as the estimation result of the user.
  • the advertisement selection unit 1611 to which the user attribute is input reads the advertisement information corresponding to the estimated user attribute from the advertisement storage unit 1621 and inputs this advertisement information to the content distribution unit 1614 (step S1809).
  • the content distribution unit 1614 reads the content requested by the user from the content storage unit 1622, adds the advertisement information to the content, and distributes it to the user terminal 1401 (step S1810).
  • the content display unit 1512 receives and displays the content including the advertisement information (step S1811).
  • FIG. 19 is an explanatory diagram showing an example of content 1901 including advertisement information 1902 distributed to the user terminal 1401 by the operation of FIG.
  • a link (URL) to another content corresponding to the user attribute may be added to the content and distributed.
  • the content distribution system according to the fourth embodiment described above has a page configuration in which the web server corresponds to the user attribute according to the estimated user attribute. This is an embodiment in which content is structured and distributed.
  • the web server includes a page configuration selection unit 2012 that selects a page configuration of content based on a user attribute instead of the advertisement selection unit. Then, the content distribution means 1614 configures the content with this page configuration and distributes it to the user terminal.
  • the present embodiment can distribute the content to the user terminal with a page configuration that is suitable for the calculated attribute of each user.
  • a page configuration that is suitable for the calculated attribute of each user.
  • a plurality of layout patterns prepared in advance correspond to the user's attributes as a method for configuring the content page.
  • the page configuration is specified, and the content is configured and distributed by the page configuration.
  • the attribute estimation apparatus 500 is used to estimate the user's attribute.
  • the estimated user attributes are mainly gender, age, occupation, and the like. This is because it is more effective to use a page structure based on a design, layout, etc. that are more acceptable depending on these attributes.
  • connection configuration of each device according to the fifth embodiment of the present invention is the same as the connection configuration according to the fourth embodiment described in FIG.
  • the web server 1403 is replaced with a web server 1403b described later.
  • the configuration of the user terminal 1401 is the same as the configuration described in FIG.
  • FIG. 20 is an explanatory diagram showing the configuration of the web server 1403b according to the fifth embodiment of the present invention.
  • the web server 1403b is different from the web server 1403 shown in FIG. 16 in that the page configuration selection unit 2012 is executed by the CPU 1602 instead of the advertisement selection unit 1611 and the advertisement storage unit 1621 is omitted in the HDD 1604. Is the difference. Except this point, the same operation as the web server 1403 shown in FIG. 16 is performed.
  • the web server 1403b specifies a page configuration corresponding to the user's attribute from a plurality of patterns prepared in advance as the content page configuration method, and the content is determined by the page configuration.
  • the content distribution unit 1614 distributes the content requested by the user with a page configuration suitable for the user attribute.
  • the page configuration selection unit 2012 receives the user attribute estimation result from the attribute estimation apparatus 500 from the input unit 1609, extracts the page configuration corresponding to the user attribute from the content storage unit 1603, and passes it to the content distribution unit 1614.
  • the operation is divided into two processes, that is, an attribute model learning process and a process of delivering content having a page configuration corresponding to the attribute estimation result.
  • the attribute model learning process is the same as the operation according to the fourth embodiment described with reference to FIG.
  • FIG. 21 is a sequence diagram illustrating an operation of a process of delivering content having a page configuration corresponding to the attribute estimation result performed by the user terminal 1401, the web server 1403b, and the attribute estimation apparatus 500 illustrated in FIGS. is there.
  • Steps S1801 to 1808 are the same as those according to the fourth embodiment described with reference to FIG. With this operation, the likelihood calculation means 510 estimates the user's attributes and is completed until it is output to the page configuration selection means 2012 through the input means 1601.
  • the page configuration selection unit 2012 to which the user attribute is input reads out the page configuration corresponding to the estimated user attribute from the content storage unit 1603 with the content received from the user (step S2109).
  • 22 and 23 are explanatory diagrams illustrating examples of two contents 2201 and 2301 having the same contents but different page configurations.
  • the page configuration selection unit 2012 inputs the read page configuration to the content distribution unit 1614.
  • the content distribution unit 1614 reads the content requested by the user from the content storage unit 1622, configures it with the input page configuration, and distributes it to the user terminal 1401 (step S2110).
  • the content display unit 1512 receives and displays this content (step S1811).
  • a pattern corresponding to the user's attribute is specified from a plurality of patterns prepared in advance, and the content is configured with the pattern. Can be delivered.
  • the search word history can also be used for user attribute estimation.
  • a likelihood calculation apparatus (spoofing detection apparatus 2400) according to the sixth embodiment of the present invention is a likelihood calculation apparatus according to the first and second embodiments, in which the input unit 101 uses a user input command as a symbol, The time interval from the time when the user inputs one input command to the next time input command is accepted as the transition time interval. And this apparatus has the impersonation detection means 2402 which judges whether a user is impersonating based on the calculated likelihood.
  • the present embodiment can detect an impersonation act by misusing this password even if the password is stolen by an unauthorized method. Hereinafter, this will be described in more detail.
  • the likelihood calculation devices according to the first and second embodiments described above are used as an impersonation detection device that impersonates a user and detects “spoofing” that enters the system. This is an embodiment.
  • “Spoofing” here means that someone who does not have the authority to operate the system intrudes into the system and steals confidential information that cannot be seen by anyone other than those who have authority. Or to do. In many cases, impersonation is performed by stealing an authorized person's ID and password in an unauthorized manner.
  • FIG. 24 is an explanatory diagram showing a configuration of an impersonation detection device 2400 according to the sixth embodiment of the present invention.
  • the impersonation detection device 240 is a computer device similar to the likelihood calculation device 300 according to the second embodiment shown in FIG. 3, and has the same input means 101, CPU 102, RAM 103, HDD 104, output as the likelihood calculation device 300. Means 105 is provided.
  • the functions and operations of the operation units called with the same names and reference numbers are the same as those described in the first to third embodiments.
  • the spoofing detection unit 2402 in addition to the likelihood calculation unit 110 and the parameter estimation unit 320 is constructed on software by the CPU 102 executing the computer program.
  • a data string (X, T) composed of N symbols X and N transition time intervals T is used.
  • FIG. 25 is a table showing an example of a data string (X, T) input to the impersonation detection device 2400 shown in FIG.
  • a column X of N symbols is a command history having a Linux (registered trademark) command as one symbol.
  • the transition time interval column T is a column of time until a next command is input after a certain command is input.
  • the likelihood calculation means 103 calculates the likelihood P (X, ⁇ ) for the model for each data string (X, T) input by the input means 101 and passes the calculation result to the output means 105.
  • the parameter estimation unit 320 estimates the model parameter 111 from the data string (X, T) input from the input unit 101, and stores the parameter 111 in the HDD 104.
  • the HDD 104 stores parameters 111 necessary for the likelihood calculation means.
  • Impersonation detection means 2402 determines whether or not the input sequence is impersonation based on the likelihood calculated by the likelihood calculation means 103.
  • the output unit 105 outputs the determination result by the spoofing detection unit 2402.
  • FIG. 26 is a flowchart showing processing of likelihood calculation and spoofing detection performed by the likelihood calculation unit 103 and the spoofing detection unit 2402 shown in FIG.
  • Steps S201 to S203 are the same as the likelihood calculation operation shown in FIG.
  • the likelihood calculated by the likelihood calculating unit 103 in this operation is output to the spoofing detecting unit 2402.
  • the impersonation detection unit 2402 determines whether or not the input sequence is impersonation from the likelihood (step S2601), and outputs the determination result via the output unit 105 (step S2602).
  • the operation in step S2601 can be determined to be spoofing when, for example, the likelihood is greater than or equal to a threshold value.
  • FIG. 27 is an explanatory diagram showing an example of determination data output via the output unit 105 in the process shown in step S2602 of FIG.
  • the threshold value is preset as “0.8”
  • the first data string is spoofed because the likelihood is “0.9”, which is higher than the threshold value. Since the likelihood of the data string is lower than the threshold of “0.3”, it is determined that the data string is not impersonated.
  • the operation of the parameter estimation means 320 estimating the parameter 111 is the same as the operation according to the second embodiment shown in FIG.
  • the configuration of the spoofing detection device 2400 may be the configuration shown in FIG. 24 excluding the parameter estimation unit 320.
  • FIG. 28 is an explanatory diagram illustrating a configuration of an impersonation detection device 2400b according to the first modification of the sixth embodiment.
  • the parameter 111 is not estimated but is stored in the HDD 104 in advance.
  • the command time interval T may be a discrete value in the data string (X, T) input to the impersonation detection device 2400.
  • FIG. 29 is an explanatory diagram showing an example of a data string (X, T) in which the command time interval T is a discrete value.
  • the command time interval T is discretized in three stages of LOW, HIGH, and MID.
  • the input data X may be a key typing history for each character instead of for each command.
  • FIG. 30 is an explanatory diagram illustrating an example of a data string (X, T) in which the input data X is a key typing history.
  • FIG. 31 is an explanatory diagram showing an example of a data string (X, T) in which the input data X is a key typing history and the command time interval T is a discrete value of three levels of LOW, HIGH, and MID.
  • the likelihood calculation apparatus (motion recognition apparatus 3200) according to the seventh embodiment of the present invention is a likelihood calculation apparatus according to the first and second embodiments, and the input unit 101 is a moving image in which a person is shown. Each input is accepted with a feature vector indicating the feature of each human image as a symbol and a time interval until the feature vector transitions from one state to the next state as a transition time interval. And this apparatus has the action recognition means 3202 which judges whether the person reflected in the person image is moving based on the calculated likelihood.
  • the present embodiment can determine whether or not a person shown in the image is moving. Hereinafter, this will be described in more detail.
  • the likelihood calculation device is used to determine the posture of a person from a moving image in which person images arranged in the image are arranged in time order. It is embodiment using it as an operation
  • movement recognition apparatus which recognizes.
  • the motion recognition device here outputs a recognition result that a person is walking when a moving image of a person walking is input, and outputs a recognition result that a person is not walking when a moving image of a person who is not walking is input. To do.
  • FIG. 32 is an explanatory diagram showing the configuration of the motion recognition apparatus 3200 according to the seventh embodiment of the present invention.
  • the motion recognition device 3200 is a computer device similar to the likelihood calculation device 300 according to the second embodiment shown in FIG. 3, and has the same input means 101, CPU 102, RAM 103, HDD 104, output as the likelihood calculation device 300. Means 105 is provided.
  • the functions and operations of the operation units called with the same names and reference numbers are the same as those described in the first to third embodiments.
  • the motion recognition means 3202 is constructed on software by the CPU 102 executing the computer program.
  • a data string (X, T) composed of N symbols X and N transition time intervals T is used.
  • a moving image is a data sequence in which person images are arranged
  • a symbol X column in the data sequence is an X ′, Y, Z axis (X ′, Y, Z) column of the person image.
  • a feature vector (X ′, Y, Z) is obtained by combining x ′, y, z of all contours into a vector having x ′, y, z at one point of the contour portion of the human region as elements.
  • FIG. 33 is an explanatory diagram showing an example of a feature vector input to the motion recognition device 3200 shown in FIG.
  • the data sequence X is a sequence of feature vectors.
  • the transition time interval T represents a time interval for transition from one feature vector to the next feature vector.
  • the feature vectors (X ′, Y, Z) acquired by the imaging device are arranged in order of time, if the difference between the feature vectors is greater than a certain threshold value between the two times, the feature vectors (X ′, Y, Z) is saved, and if it is small, it is not deleted and saved. Therefore, the time interval for transition from one feature vector to the next feature vector represents the feature of the interval at which the person's movement has changed significantly.
  • FIG. 34 is an explanatory diagram showing an example of input data input to the motion recognition device 3200 shown in FIG.
  • a plurality of input data may be prepared as shown in FIG.
  • M data strings (X, T) are input.
  • the likelihood calculating means 103 calculates the likelihood P (X, ⁇ ) for the model for each data string (X, T) input by the input means 101 using the model parameter 111, and the calculation result Is transferred to the motion recognition means 3202.
  • the parameter estimation unit 320 estimates the model parameter 111 from the M data strings (X, T) input from the input unit 101 and passes the estimation result to the HDD 104.
  • the HDD 104 stores the parameter 111.
  • the motion recognition unit 3202 determines whether or not the input sequence is a motion that applies to the model based on the likelihood calculated by the likelihood calculation unit 103.
  • the output unit 105 outputs the determination result by the motion recognition unit 3202.
  • FIG. 35 is a flowchart showing likelihood calculation and action recognition processing performed by the likelihood calculation means 103 and the action recognition means 3202 shown in FIG.
  • Steps S201 to S203 are the same as the likelihood calculation operation shown in FIG.
  • the likelihood calculated by the likelihood calculating unit 103 in this operation is output to the operation recognizing unit 3202.
  • the motion recognition unit 3202 determines whether or not the input sequence is a motion that applies to the model from the likelihood (step S3501), and outputs the determination result via the output unit 105 (step S3502).
  • the operation in step S3501 can be determined to be an operation that applies to the model when the likelihood is equal to or greater than a threshold, for example.
  • FIG. 36 is an explanatory diagram illustrating an example of determination data output via the output unit 105 in the operation of step S3502 of FIG.
  • a threshold value is preset as “0.8” for the walking motion model, and the first data string has a likelihood “0.9” which is higher than the threshold value. Therefore, it is applied to the model representing walking, and the second data string is lower than the threshold value with a likelihood “0.3”, so it is determined that it does not apply to the model.
  • the operation in which the parameter estimation unit 320 estimates the parameter 111 for the data sequence in which a certain person performs one operation is the same as the operation according to the second embodiment shown in FIG.
  • M data strings X and T generated from a moving image in which a person walks are input, and model parameters are estimated. Thereafter, when certain data strings X and T are input, it is possible to recognize whether each data string is walking or not.
  • FIG. 37 is an explanatory diagram illustrating a configuration of an action recognition device 3200b according to Modification 1 of the seventh embodiment.
  • the parameter 111 is not estimated but is stored in the HDD 104 in advance.
  • the feature vector time interval T can be a discrete value in the data string (X, T) input to the motion recognition device 3200 or 3200b.
  • the motion recognition device 3200 may extract feature vectors from the joint angles of a person acquired by motion capture or the like instead of extracting feature vectors from a person image taken by a camera.
  • the motion recognition device 3200 instead of characterizing the coordinates X, Y, and Z of the image, Roll, Pitch, and Yaw indicating the rotational motion angle around each axis are used.
  • FIG. 38 is an explanatory diagram showing an example of a data string (X, T) using the Roll, Pitch, and Yaw.
  • the feature vector time interval T can be a discrete value as in the second modification.
  • a user attribute estimation device examples of a user attribute estimation device, an impersonation detection device, and an action recognition device in the content distribution system are shown in the embodiment.
  • the present invention can be widely applied to the use of calculating the likelihood for a specific statistical model for a data series expressed by a combination of a symbol and a transition time interval.

Abstract

Provided is a likelihood estimation device, etc., for a data series expressed by a combination of a symbol and a transition time gap, wherein the transition time gap is taken into account. The likelihood estimation device (100) is provided with an input means (101) to receive input of the data series expressed by the combination of the symbol and the transition time gap; a storage means (104) to store a parameter (111) representing a specific statistical model of the data series; a likelihood calculation means (110) to calculate the likelihood for the statistical model of the data series using the parameter; and an output means (105) to output the likelihood calculated by the likelihood calculation means.

Description

尤度推定装置、コンテンツ配信システム、尤度推定方法および尤度推定プログラムLikelihood estimation apparatus, content distribution system, likelihood estimation method, and likelihood estimation program
　本発明は隠れマルコフモデルなどで表されるデータモデルに対するデータ列の尤度推定の技術に関し、特にシンボルと遷移時間間隔との組み合わせによって表現されるデータ系列に対する、その遷移時間間隔も考慮した尤度推定の技術に関する。 The present invention relates to a technique for estimating the likelihood of a data string for a data model represented by a hidden Markov model, and more particularly to a data sequence expressed by a combination of a symbol and a transition time interval in consideration of the transition time interval. It relates to estimation technology.
　あるデータ列とその他のデータ列との類似を測る技術がある。さらに、データ列から頻出シーケンスを抽出するなどの技術がある。ここでいうデータ列（またはシーケンスデータ）とは、例えばキーボードによって文字を入力すると、入力した文字がデータであり、入力した文字が並んだものがデータ列となる。 There is a technique to measure the similarity between a certain data string and other data strings. Furthermore, there is a technique such as extracting a frequent sequence from a data string. As used herein, the data string (or sequence data) is, for example, when characters are input using a keyboard, the input characters are data, and the input characters are arranged in a data string.
　データ列を扱う方法は、大きく２つに分かれる。１つは、数え上げる方法である。データ列間の類似度は、同じデータがどれだけ含まれていて、どの順番に並んでいるかを数え上げて計算する。さらに、データ列全体で頻出する部分的なシーケンス（特定の文字列など）を数え上げて高速に抽出する。もう１つは、統計モデルを用いてデータ列を扱う方法である。例えば、データ列が、ある確率分布に従うと仮定する方法である。したがって、データ列間の距離は、確率分布の尤度で測ることができる。 方法 There are two main methods for handling data strings. One is a counting method. The similarity between the data strings is calculated by counting how many pieces of the same data are included and in what order. Furthermore, partial sequences (such as specific character strings) that frequently appear in the entire data string are counted and extracted at high speed. The other is a method of handling a data string using a statistical model. For example, this is a method of assuming that the data string follows a certain probability distribution. Therefore, the distance between the data strings can be measured by the probability distribution probability.
　ここで、尤度とは、対象としている確率分布から発生しやすい度合いを表す。一般にはデータ列のデータのみを入力としているが、本来データ列には、データを取得する時間情報が付与されている。時間情報の付いたデータ列にはデータと次のデータが観測されるまでの時間間隔という特徴が含まれているので、この時間間隔を分析に使用することができる。 Here, the likelihood represents a degree that is likely to occur from the target probability distribution. In general, only data in a data string is input, but originally the data string is given time information for acquiring data. Since the data string with time information includes a feature of time interval until data and the next data are observed, this time interval can be used for analysis.
　ここで、データ列の中であるデータから次のデータが観測されるまでの時間間隔を遷移時間間隔と呼ぶ。データ列の遷移時間間隔も考慮してデータ列を分析すると、データには特徴が表れなくても、遷移時間間隔に特徴が表れる場合がある。 Here, the time interval from the data in the data string until the next data is observed is called the transition time interval. If the data string is analyzed in consideration of the transition time interval of the data string, the characteristic may appear in the transition time interval even if no characteristic appears in the data.
　たとえば、ウェブコンテンツを配信するシステムを例に取ると、各々のユーザは自分の興味のある事項に関するコンテンツは長時間かけてじっくり読むのに対し、興味のない事項に関するコンテンツは短時間で読み飛ばそうとする。従って、ユーザが過去にどのコンテンツをどの順番で訪問したのかというコンテンツ訪問履歴をデータとし、これに各コンテンツに滞在した時間を示す遷移時間間隔を合わせたデータ列から、各ユーザがどのような事項について興味があるかをより明確に見出せることが期待できる。 For example, in the case of a system that distributes web content, each user will read content related to his / her interests over a long period of time, while content related to uninterested items will be skipped in a short time. And Therefore, what is the content of each user from the data string that includes the content visit history that the user has visited in what order in the past and the transition time interval indicating the time spent in each content. It can be expected to find more clearly whether you are interested in.
　各ユーザが興味のある事項について知ることができれば、コンテンツに各ユーザの興味の対象に合わせた内容の追加コンテンツ（広告など）を組み合わせて配信することができる。これによってより有効な追加コンテンツを配信できれば、広告費収入の向上などのようなメリットを得られることが期待できる。 If it is possible to know the matters that each user is interested in, the content can be delivered in combination with additional content (such as advertisements) that is tailored to each user's interest. If more effective additional content can be distributed in this way, it can be expected that benefits such as an improvement in advertising revenue will be obtained.
　また、コンピュータにログインする際のパスワード入力を例に取ると、正規のＩＤおよびパスワードを持つユーザＡは、普段からパスワード入力に慣れているので、素早く正確にパスワードを打ち込むことができる。ここで、ユーザＡのパスワードを何らかの不正な手段（たとえばキーロガー、スニファリングなど）で入手したユーザＢが、そのパスワードを入力して、ユーザＡになりすましてそのコンピュータにログインしようとしている。その場合、ユーザＢは、ユーザＡのパスワードの入力には慣れていないため、このパスワードを素早く正確に打ち込むことが難しい。 Also, taking the password input when logging in to the computer as an example, user A who has a regular ID and password is accustomed to entering the password on a regular basis, and can therefore type in the password quickly and accurately. Here, the user B who has obtained the password of the user A by some unauthorized means (for example, key logger, sniffing, etc.) enters the password, impersonates the user A and attempts to log in to the computer. In that case, since the user B is not used to inputting the password of the user A, it is difficult to input the password quickly and accurately.
　この場合、ユーザＡとユーザＡになりすましたユーザＢは同一のパスワードを入力するので、入力されたパスワードのデータ列だけではなりすましを検出できない。しかし、ユーザＡとユーザＢとではパスワードを入力するスピードが明らかに異なる。そこで、各コマンドまたは各入力文字をデータとし、このデータ間の遷移時間間隔を調べることによって、なりすましを有効に検出できることが期待できる。 In this case, since user B pretending to be user A and user A inputs the same password, impersonation cannot be detected only by the data string of the input password. However, the speed of inputting a password is clearly different between user A and user B. Therefore, it can be expected that impersonation can be detected effectively by using each command or each input character as data and examining the transition time interval between the data.
　以上で示した各々の例のように、ユーザの操作に係る操作履歴情報をデータ列とすると、このデータ列のデータそのものだけでは表れない特徴がデータ間の遷移時間間隔に表れることがあるので、遷移時間間隔を分析することには意義がある。 As in each example shown above, if the operation history information related to the user's operation is a data string, a feature that does not appear only by the data itself of this data string may appear in the transition time interval between the data. It is meaningful to analyze the transition time interval.
　以下、非特許文献１および非特許文献２に記載された、統計モデルによってデータ列を分析する方法について説明する。データ列を扱う統計モデルの中で、最も応用されているものの１つに、マルコフモデルや隠れマルコフモデルがある。 Hereinafter, a method for analyzing a data string using a statistical model described in Non-Patent Document 1 and Non-Patent Document 2 will be described. One of the most applied statistical models for handling data strings is a Markov model or a hidden Markov model.
　隠れマルコフモデル（Hidden Markov Model、以後ＨＭＭという）は、データ列をデータの発生モデルをもつ隠れ状態の遷移として表現するデータモデルである。実際には観測できないデータ列であるという意味で、「隠れ」状態と呼ばれる。実際に観測されない隠れ状態が遷移することで、データ列が生成されると仮定したモデルである。ＨＭＭは、データをシンボル（データの値）として表現するものであり、文字認識、自然言語処理、音声認識、動作認識などの分野で利用されている。 The hidden Markov model (Hidden-Markov Model, hereinafter referred to as HMM) is a data model that expresses a data string as a hidden state transition having a data generation model. It is called a “hidden” state in the sense that it is a data string that cannot actually be observed. This model assumes that a data string is generated when a hidden state that is not actually observed transitions. The HMM expresses data as symbols (data values), and is used in fields such as character recognition, natural language processing, speech recognition, and motion recognition.
　以後、本明細書の数式以外の行では、たとえば文字Ｘに下付き添字１が付いた状態をＸ１、文字Ｘに上付き添字２が付いた状態をＸ（２）、文字Ｘに下付き添字１と上付き添字２が同時に付いた状態をＸ１（２）などのような形で表記するものとする。 Thereafter, in the lines other than the mathematical expressions in the present specification, for example, the state where the subscript 1 is attached to the character X is X1, the state where the subscript 2 is attached to the character X is X (2), and the subscript is attached to the character X. A state in which 1 and superscript 2 are added at the same time is expressed in a form such as X1 (2).
　非特許文献１および非特許文献２に記載されたＨＭＭでデータ列をモデル化する統計モデルを数１に示す。入力データ列をＸとし、データＸの長さをＮとする。Ｋはデータ列ＸのＨＭＭの混合数であり、データ列Ｘが混合しない場合はＫ＝１となる。Ｘ＝｛Ｘ１，Ｘ２，…，Ｘｎ｝とする。
Figure JPOXMLDOC01-appb-M000001
 A statistical model for modeling a data string by the HMM described in Non-Patent Document 1 and Non-Patent Document 2 is shown in Equation 1. The input data string is X, and the length of the data X is N. K is the number of HMMs mixed in the data string X. When the data string X is not mixed, K = 1. Let X = {X1, X2,..., Xn}.
Figure JPOXMLDOC01-appb-M000001
　ここで、ＩをＨＭＭの隠れ状態数とすると、数２が成立する。Ｓ＝｛Ｓ１，Ｓ２，…，Ｓｎ｝とすると、Ｓ１，Ｓ２，…，Ｓｎは、Ｓに値をとる確率変数とする。πｋは、各コンポーネントの混合確率を表す。Ｐ（Ｘ｜θｋ）は、コンポーネントＫのＨＭＭを表わす。θｋ＝｛πｋ，Γｋ，Ａｋ，Ｂｋ｝は、コンポーネントＫに関するパラメータである。Γｋは、コンポーネントｋ番目の隠れ変数の初期確率のパラメータを表す。Ａｋは、コンポーネントｋ番目の隠れ変数の遷移確率のパラメータを表す。Ｂｋは、コンポーネントｋ番目のシンボルの生起確率のパラメータを表す。
Figure JPOXMLDOC01-appb-M000002
 Here, when I is the number of HMM hidden states, Equation 2 is established. Assuming that S = {S1, S2,..., Sn}, S1, S2,. πk represents the mixing probability of each component. P (X | θk) represents the HMM of the component K. θk = {πk, Γk, Ak, Bk} is a parameter related to the component K. Γk represents the parameter of the initial probability of the component kth hidden variable. Ak represents a parameter of the transition probability of the component kth hidden variable. Bk represents an occurrence probability parameter of the k-th component symbol.
Figure JPOXMLDOC01-appb-M000002
　一方、特許文献１には、ＨＭＭによる類似度から動画像の類似度を判定して検索を行うという技術が記載されている。特許文献２には、顧客の購買行動をＨＭＭによってモデル化するという技術が記載されている。特許文献３には、人物画像データを入力データとして、ＨＭＭにより人物が転倒したか否かを検出するという技術が記載されている。 On the other hand, Patent Document 1 describes a technique of performing a search by determining the similarity of moving images from the similarity of HMM. Patent Document 2 describes a technique of modeling customer purchase behavior using an HMM. Patent Document 3 describes a technique of detecting whether a person has fallen by HMM using human image data as input data.
　特許文献４には、データの一部が欠落している不完全データからＥＭアルゴリズムによってパラメータを推定する最尤推定に関する技術が記載されている。また特許文献５には、無線装置を搭載した位置センサ、接触センサ、オーディオセンサ、コンパスセンサ等から取得できるセンサ情報を入力すると、ＨＭＭによって状況を認識して、この状況に応じたウェブサービスを提供するという技術が記載されている。 Patent Document 4 describes a technique relating to maximum likelihood estimation in which parameters are estimated by an EM algorithm from incomplete data in which a part of data is missing. In Patent Document 5, when sensor information that can be acquired from a position sensor, a contact sensor, an audio sensor, a compass sensor, or the like equipped with a wireless device is input, the situation is recognized by the HMM and a web service corresponding to the situation is provided. The technology to do is described.
特開２００６－１７８９７４号公報JP 2006-178974 A 特開２００８－１５２３２１号公報JP 2008-152321 A 特開２００８－１５２７１７号公報JP 2008-152717 A 特開平０９－０６２６５２号公報JP 09-062652 A 特表２００４－５３５０００号公報Special table 2004-535000 gazette
　前述のウェブコンテンツの配信やなりすましの検出などの用途には、シンボルと遷移時間間隔との組み合わせを考慮した尤度推定が必要であるが、そのような尤度推定を可能とする技術は存在しなかった。非特許文献１および非特許文献２には、数え上げることでデータ列を分析する方法で、データだけを用いる分析の改良として、遷移時間間隔を考慮した技術が記載されている。しかしながら、これらの文献に記載の統計モデルによってデータ列を分析する技術では、データ（シンボル）とは別に遷移時間間隔をモデルに組み込むことができない。 For applications such as web content delivery and spoofing detection, it is necessary to estimate the likelihood in consideration of the combination of the symbol and the transition time interval. However, there is a technology that enables such likelihood estimation. There wasn't. Non-Patent Document 1 and Non-Patent Document 2 describe a technique that considers transition time intervals as an improvement of analysis using only data in a method of analyzing a data string by counting. However, in the technique of analyzing a data string using a statistical model described in these documents, a transition time interval cannot be incorporated into the model separately from data (symbols).
　非特許文献１および非特許文献２に記載の技術で、χｎを多次元と考えて、ｎ－１個目のシンボルからｎ個目のシンボルへの遷移時間間隔をχｎの要素に含めることで、前述の数１に遷移時間間隔を組み込むことができる。しかし、遷移時間間隔をシンボルと切り離してモデル化することはできない。さらに、この方法だとシンボルと遷移時間間隔のペアの組み合わせが多くなるため、計算時間やデータ容量が増大し、現実的ではない。 In the technique described in Non-Patent Document 1 and Non-Patent Document 2, χn is considered multidimensional, and the transition time interval from the (n−1) th symbol to the nth symbol is included in the χn element, The transition time interval can be incorporated into Equation 1 above. However, the transition time interval cannot be modeled separately from the symbol. Furthermore, since this method increases the number of combinations of symbol and transition time interval pairs, calculation time and data capacity increase, which is not practical.
　また、特許文献１および特許文献２に記載の技術は、ＨＭＭを用いて尤度の計算を行っているが、上記で問題としているシンボルと遷移時間間隔との組み合わせによって表現されるデータ系列について扱っているものではない。特許文献３および特許文献５に記載の技術は、ＨＭＭによってデータ列を解析しているが、遷移時間間隔を用いていない。特許文献４は、ＥＭアルゴリズムによるパラメータの推定の技術について述べているに過ぎない。 In addition, the techniques described in Patent Document 1 and Patent Document 2 calculate likelihood using an HMM, but deal with a data series expressed by a combination of a symbol and a transition time interval in question above. It is not what you have. The techniques described in Patent Document 3 and Patent Document 5 analyze the data string by HMM, but do not use the transition time interval. Patent Document 4 only describes a parameter estimation technique using an EM algorithm.
　本発明の目的は、ユーザの操作に係る操作履歴情報に基づいて、当該ユーザの属性を有効に判別する手段としての尤度推定を可能とする尤度推定装置、コンテンツ配信システム、尤度推定方法および尤度推定プログラムを提供することにある。 An object of the present invention is to provide a likelihood estimation device, a content distribution system, and a likelihood estimation method that enable likelihood estimation as means for effectively discriminating a user attribute based on operation history information related to a user operation. And providing a likelihood estimation program.
　上記目的を達成するため、本発明に係る尤度推定装置は、シンボルと遷移時間間隔との組み合わせによって表現されるデータ系列の入力を受け付ける入力手段と、データ系列の特定の統計モデルを表すパラメータを記憶する記憶手段と、パラメータを用いてデータ系列の統計モデルに対する尤度を計算する尤度計算手段と、この尤度計算手段によって計算された尤度を出力する出力手段を備えたことを特徴とする。 In order to achieve the above object, a likelihood estimation apparatus according to the present invention includes an input unit that receives an input of a data sequence expressed by a combination of a symbol and a transition time interval, and a parameter that represents a specific statistical model of the data sequence. A storage means for storing, a likelihood calculation means for calculating a likelihood for a statistical model of a data series using parameters, and an output means for outputting the likelihood calculated by the likelihood calculation means, To do.
　上記目的を達成するため、本発明に係るコンテンツ配信システムは、ユーザが操作するユーザ端末と、ウェブサーバと、尤度計算装置とがネットワークを介して相互に接続されたコンテンツ配信システムであって、ウェブサーバが、ユーザ端末からのリクエストを受け付けるリクエスト受付手段と、リクエストに対応したコンテンツをユーザ端末に配信するコンテンツ配信手段と、ユーザにリクエストされたコンテンツの履歴を示すコンテンツ訪問履歴と、ユーザが一つのコンテンツから次のコンテンツを訪問するまでのコンテンツ訪問間隔とを記憶するユーザリクエスト記憶部と、コンテンツ訪問履歴をシンボル、コンテンツ訪問間隔を遷移時間間隔とするデータ系列として尤度計算装置に入力する出力手段とを有し、尤度計算装置が、シンボルと遷移時間間隔との組み合わせによって表現されるデータ系列の入力を受け付ける入力手段と、データ系列の特定の統計モデルを表すパラメータを記憶する記憶手段と、パラメータを用いてデータ系列の統計モデルに対する尤度を計算する尤度計算手段と、この尤度計算手段によって計算された尤度を出力する出力手段を備え、ウェブサーバが、尤度計算装置で計算されたユーザの属性の入力を受け付ける入力手段と、ユーザの属性に基づいてコンテンツに追加する広告を選択する広告選択手段とを有し、コンテンツ配信手段がコンテンツに広告を追加してユーザ端末に配信する機能を持つことを特徴とする。 In order to achieve the above object, a content distribution system according to the present invention is a content distribution system in which a user terminal operated by a user, a web server, and a likelihood calculation device are connected to each other via a network. A web server receives a request from a user terminal, a content receiving unit that distributes content corresponding to the request to the user terminal, a content visit history that indicates a history of the content requested by the user, and a user A user request storage unit that stores a content visit interval from one content to a visit to the next content, and an output that is input to the likelihood calculation device as a data sequence in which the content visit history is a symbol and the content visit interval is a transition time interval A likelihood calculating device, Input means for receiving an input of a data series expressed by a combination of a symbol and a transition time interval, a storage means for storing a parameter representing a specific statistical model of the data series, and a likelihood for the statistical model of the data series using the parameter. Input means for receiving a user attribute input calculated by the likelihood calculation device, comprising: a likelihood calculation means for calculating the degree; and an output means for outputting the likelihood calculated by the likelihood calculation means. And an advertisement selection means for selecting an advertisement to be added to the content based on the user attribute, and the content distribution means has a function of adding an advertisement to the content and delivering it to the user terminal.
　上記目的を達成するため、本発明に係る他のコンテンツ配信システムは、ユーザが操作するユーザ端末と、ウェブサーバと、尤度計算装置とがネットワークを介して相互に接続されたコンテンツ配信システムであって、ウェブサーバが、ユーザ端末からのリクエストを受け付けるリクエスト受付手段と、リクエストに対応したコンテンツをユーザ端末に配信するコンテンツ配信手段と、ユーザにリクエストされたコンテンツの履歴を示すコンテンツ訪問履歴と、ユーザが一つのコンテンツから次のコンテンツを訪問するまでのコンテンツ訪問間隔とを記憶するユーザリクエスト記憶部と、コンテンツ訪問履歴をシンボル、コンテンツ訪問間隔を遷移時間間隔とするデータ系列として尤度計算装置に入力する出力手段とを有し、尤度計算装置が、シンボルと遷移時間間隔との組み合わせによって表現されるデータ系列の入力を受け付ける入力手段と、データ系列の特定の統計モデルを表すパラメータを記憶する記憶手段と、パラメータを用いてデータ系列の統計モデルに対する尤度を計算する尤度計算手段とを有すると共に、この尤度計算手段によって計算された尤度を出力する出力手段を備え、ウェブサーバが、尤度計算装置で計算されたユーザの属性の入力を受け付ける入力手段と、ユーザの属性に基づいてコンテンツのページ構成を選択するページ構成選択手段とを有し、コンテンツ配信手段がページ構成によってコンテンツをユーザ端末に配信する機能を持つことを特徴とする。 In order to achieve the above object, another content distribution system according to the present invention is a content distribution system in which a user terminal operated by a user, a web server, and a likelihood calculation device are connected to each other via a network. The web server receives a request from the user terminal, a request accepting unit, a content distributing unit that distributes the content corresponding to the request to the user terminal, a content visit history indicating a history of the content requested by the user, and a user A user request storage unit that stores a content visit interval from one content to a visit to the next content, and a data sequence having a content visit history as a symbol and a content visit interval as a transition time interval are input to the likelihood calculation device. A likelihood calculation device An input means for receiving an input of a data series expressed by a combination of a symbol and a transition time interval, a storage means for storing a parameter representing a specific statistical model of the data series, and a statistical model of the data series using the parameter A likelihood calculating means for calculating likelihood, and an output means for outputting the likelihood calculated by the likelihood calculating means, wherein the web server inputs the attribute of the user calculated by the likelihood calculating device. And a page configuration selection unit that selects a page configuration of content based on a user attribute, and the content distribution unit has a function of distributing content to a user terminal by the page configuration .
　上記目的を達成するため、本発明に係る尤度計算方法は、シンボルと遷移時間間隔との組み合わせによって表現されるデータ系列に対して、データ系列の特定の統計モデルに対する尤度を計算する尤度計算方法であって、入力手段がデータ系列の入力を受け付け、記憶手段から統計モデルのパラメータを読み出し、パラメータを用いて尤度計算手段がデータ系列の統計モデルに対する尤度を計算し、計算された尤度を出力手段に出力することを特徴とする。 In order to achieve the above object, a likelihood calculation method according to the present invention is a likelihood of calculating a likelihood for a specific statistical model of a data sequence for a data sequence expressed by a combination of a symbol and a transition time interval. A calculation method in which an input unit receives an input of a data series, reads parameters of a statistical model from a storage unit, and a likelihood calculation unit calculates a likelihood for the statistical model of the data series using the parameters, and is calculated The likelihood is output to the output means.
　上記目的を達成するため、本発明に係る尤度計算プログラムは、シンボルと遷移時間間隔との組み合わせによって表現されるデータ系列に対して、データ系列の特定の統計モデルに対する尤度を計算する尤度計算装置であって、データ系列の入力を受け付ける手順と、統計モデルのパラメータを読み出す手順と、パラメータを用いてデータ系列の統計モデルに対する尤度を計算する手順と、計算された尤度を出力手段に出力する手順とをコンピュータに実行させることを特徴とする。 In order to achieve the above object, a likelihood calculation program according to the present invention calculates a likelihood for a data sequence expressed by a combination of a symbol and a transition time interval, with respect to a specific statistical model of the data sequence. A calculation device, a procedure for receiving an input of a data series, a procedure for reading a parameter of a statistical model, a procedure for calculating a likelihood for the statistical model of the data series using the parameter, and a means for outputting the calculated likelihood And outputting the data to the computer.
　本発明は、上述したようにシンボルと遷移時間間隔との組み合わせによって表現されるデータ系列について、この遷移時間間隔も考慮して既知の統計モデルに対する尤度を算出するように構成したので、ユーザの操作に係る操作履歴情報に基づいて、当該ユーザの属性を有効に判別する手段としての尤度推定が可能である。 Since the present invention is configured to calculate the likelihood for a known statistical model in consideration of the transition time interval for the data series expressed by the combination of the symbol and the transition time interval as described above, the user's Based on the operation history information related to the operation, likelihood estimation as means for effectively determining the attribute of the user can be performed.
本発明の第１の実施形態に係る尤度計算装置の構成を示す説明図である。It is explanatory drawing which shows the structure of the likelihood calculation apparatus which concerns on the 1st Embodiment of this invention. 図１に示した尤度計算手段の動作を表すフローチャートである。It is a flowchart showing operation | movement of the likelihood calculation means shown in FIG. 図３は、本発明の第２の実施形態に係る尤度計算装置の構成を示す説明図である。FIG. 3 is an explanatory diagram showing the configuration of the likelihood calculating apparatus according to the second embodiment of the present invention. 図３に示したパラメータ推定手段がパラメータを推定する動作について示すフローチャートである。It is a flowchart shown about the operation | movement which the parameter estimation means shown in FIG. 3 estimates a parameter. 本発明の第３の実施形態に係る属性推定装置の構成を示す説明図である。It is explanatory drawing which shows the structure of the attribute estimation apparatus which concerns on the 3rd Embodiment of this invention. 図５で示した属性モデル学習手段が、属性ごとに学習したモデルのパラメータの例を示す説明図である。It is explanatory drawing which shows the example of the parameter of the model which the attribute model learning means shown in FIG. 5 learned for every attribute. 図５で示した尤度計算手段および属性モデル学習手段に入力手段から入力される、コンテンツ訪問履歴Ｘとコンテンツ訪問間隔Ｔの例を示す説明図である。FIG. 6 is an explanatory diagram illustrating an example of a content visit history X and a content visit interval T that are input from the input unit to the likelihood calculating unit and the attribute model learning unit illustrated in FIG. 5. 図５で示した属性モデル学習手段が行う属性モデル学習処理を表すフローチャートである。It is a flowchart showing the attribute model learning process which the attribute model learning means shown in FIG. 5 performs. 図５で示した尤度計算手段が行うユーザ属性の推定処理を表すフローチャートである。It is a flowchart showing the estimation process of the user attribute which the likelihood calculation means shown in FIG. 5 performs. 図９のステップＳ９０３で算出された、各属性に対する尤度の例を示す説明図である。It is explanatory drawing which shows the example of the likelihood with respect to each attribute calculated by step S903 of FIG. 図９のステップＳ９０５で出力された、ユーザの属性の出力の例を示す説明図である。It is explanatory drawing which shows the example of the output of a user's attribute output by step S905 of FIG. 本発明の第３の実施形態の変形例１で入力手段から入力される、コンテンツ訪問履歴Ｘとコンテンツ訪問間隔Ｔの例を示す説明図である。It is explanatory drawing which shows the example of the content visit log | history X and the content visit interval T input from the input means in the modification 1 of the 3rd Embodiment of this invention. 本発明の第３の実施形態の変形例２で入力手段から入力される、コンテンツ訪問履歴Ｘとコンテンツ訪問間隔Ｔおよび検索ワードの例を示す説明図である。It is explanatory drawing which shows the example of the content visit history X, the content visit interval T, and a search word which are input from the input means in the modification 2 of the 3rd Embodiment of this invention. 本発明の第４の実施形態に係るコンテンツ配信システムで、各装置の接続構成を示す説明図である。It is explanatory drawing which shows the connection structure of each apparatus in the content delivery system which concerns on the 4th Embodiment of this invention. 図１４で示したユーザ端末の構成を示す説明図である。It is explanatory drawing which shows the structure of the user terminal shown in FIG. 図１４で示したウェブサーバの構成を示す説明図である。It is explanatory drawing which shows the structure of the web server shown in FIG. 図１４～１６に示したユーザ端末、ウェブサーバ、および属性推定装置が行う属性モデル学習処理の動作を示すシーケンス図である。FIG. 17 is a sequence diagram showing an operation of an attribute model learning process performed by the user terminal, the web server, and the attribute estimation device shown in FIGS. 図１４～１６に示したユーザ端末、ウェブサーバ、および属性推定装置が行う、推定した属性に対応する広告の配信処理の動作の例を示すシーケンス図である。FIG. 17 is a sequence diagram illustrating an example of an operation of an advertisement distribution process corresponding to an estimated attribute performed by the user terminal, the web server, and the attribute estimation device illustrated in FIGS. 図１８の動作によってユーザ端末に配信された、広告情報を含むコンテンツの例を示す説明図である。It is explanatory drawing which shows the example of the content containing advertisement information delivered to the user terminal by the operation | movement of FIG. 本発明の第５の実施形態に係るウェブサーバの構成を示す説明図である。It is explanatory drawing which shows the structure of the web server which concerns on the 5th Embodiment of this invention. 図１４、１５、２０に示したユーザ端末、ウェブサーバ、および属性推定装置が行う、属性推定結果に対応したページ構成のコンテンツを配信する処理の動作を示すシーケンス図である。It is a sequence diagram which shows the operation | movement of the process which delivers the content of the page structure corresponding to the attribute estimation result which the user terminal, web server, and attribute estimation apparatus which were shown to FIG. 同一のコンテンツで、ページ構成が異なる２つのコンテンツの例を示す説明図である。It is explanatory drawing which shows the example of two content from which a page structure differs with the same content. 同一のコンテンツで、ページ構成が異なる２つのコンテンツの例を示す説明図である。It is explanatory drawing which shows the example of two content from which a page structure differs with the same content. 本発明の第６の実施形態に係るなりすまし検知装置の構成を示す説明図である。It is explanatory drawing which shows the structure of the impersonation detection apparatus which concerns on the 6th Embodiment of this invention. 図２４に示したなりすまし検知装置に入力されるデータ列（Ｘ、Ｔ）の例を示す表である。It is a table | surface which shows the example of the data sequence (X, T) input into the impersonation detection apparatus shown in FIG. 図２４で示した尤度計算手段およびなりすまし検知手段が行う尤度の計算およびなりすましの検知の処理を表すフローチャートである。FIG. 25 is a flowchart illustrating likelihood calculation and spoofing detection processing performed by the likelihood calculation unit and the spoofing detection unit illustrated in FIG. 24. 図２６のステップＳ２６０２に示した処理で出力手段を介して出力された判定データの例を示す説明図である。It is explanatory drawing which shows the example of the determination data output via the output means by the process shown to step S2602 of FIG. 本発明の第６の実施形態の変形例１に係るなりすまし検知装置の構成を示す説明図である。It is explanatory drawing which shows the structure of the impersonation detection apparatus which concerns on the modification 1 of the 6th Embodiment of this invention. 本発明の第６の実施形態の変形例２で、なりすまし検知装置に入力されるデータ列（Ｘ、Ｔ）の例を示す説明図である。It is explanatory drawing which shows the example of the data sequence (X, T) input into the impersonation detection apparatus in the modification 2 of the 6th Embodiment of this invention. 本発明の第６の実施形態の変形例３で、なりすまし検知装置に入力されるデータ列（Ｘ、Ｔ）の例を示す説明図である。It is explanatory drawing which shows the example of the data sequence (X, T) input into the impersonation detection apparatus in the modification 3 of the 6th Embodiment of this invention. 本発明の第６の実施形態の変形例４で、なりすまし検知装置に入力されるデータ列（Ｘ、Ｔ）の例を示す説明図である。It is explanatory drawing which shows the example of the data sequence (X, T) input into the impersonation detection apparatus in the modification 4 of the 6th Embodiment of this invention. 本発明の第７の実施形態に係る動作認識装置の構成を示す説明図である。It is explanatory drawing which shows the structure of the action recognition apparatus which concerns on the 7th Embodiment of this invention. 図３２に示した動作認識装置に入力される特長ベクトルの例を示す説明図である。It is explanatory drawing which shows the example of the feature vector input into the motion recognition apparatus shown in FIG. 図３２に示した動作認識装置に入力される入力データの例を示す説明図である。It is explanatory drawing which shows the example of the input data input into the action recognition apparatus shown in FIG. 図３２で示した尤度計算手段および動作認識手段が行う尤度の計算および動作の認識の処理を表すフローチャートである33 is a flowchart showing likelihood calculation and action recognition processing performed by the likelihood calculation means and the action recognition means shown in FIG. 32. 図３５のステップＳ３５０２の動作で、出力手段を介して出力された判定データの例を示す説明図である。It is explanatory drawing which shows the example of the determination data output via the output means by operation | movement of step S3502 of FIG. 本発明の第７の実施形態の変形例１に係るなりすまし検知装置の構成を示す説明図である。It is explanatory drawing which shows the structure of the impersonation detection apparatus which concerns on the modification 1 of the 7th Embodiment of this invention. 本発明の第７の実施形態の変形例３で、動作認識装置に入力されるＲｏｌｌ，Ｐｉｔｃｈ，Ｙａｗを用いたデータ列（Ｘ、Ｔ）の例を示す説明図である。It is explanatory drawing which shows the example of the data sequence (X, T) using Roll, Pitch, and Yaw input into the action recognition apparatus in the modification 3 of the 7th Embodiment of this invention.
（第１の実施形態）
　以下、本発明の第１の実施形態の構成について添付図１に基づいて説明する。
　最初に、本実施形態の基本的な内容について説明し、その後でより具体的な内容について説明する。
　本実施形態に係る尤度計算装置１００は、シンボルと遷移時間間隔との組み合わせによって表現されるデータ系列に対して、データ系列の特定の統計モデルに対する尤度を計算する尤度計算装置である。そして、データ系列の入力を受け付ける入力手段１０１と、統計モデルのパラメータ１１１を記憶する記憶手段（ＨＤＤ１０４）と、パラメータ１１１を用いてデータ系列の統計モデルに対する尤度を計算する尤度計算手段１１０と、計算された尤度を出力する出力手段１０５とを備える。 (First embodiment)
Hereinafter, the structure of the 1st Embodiment of this invention is demonstrated based on attached FIG.
First, the basic content of the present embodiment will be described, and then more specific content will be described.
The likelihood calculation apparatus 100 according to the present embodiment is a likelihood calculation apparatus that calculates the likelihood of a specific statistical model of a data series for a data series expressed by a combination of a symbol and a transition time interval. An input unit 101 that receives an input of a data sequence, a storage unit (HDD 104) that stores a parameter 111 of the statistical model, a likelihood calculation unit 110 that calculates a likelihood for the statistical model of the data sequence using the parameter 111, Output means 105 for outputting the calculated likelihood.
　ここで、統計モデルは隠れマルコフモデルまたはマルコフモデルのうちのいずれかによって表現され、遷移時間間隔が離散値または連続値のうちのいずれかによって表現されるものである。 Here, the statistical model is expressed by either a hidden Markov model or a Markov model, and the transition time interval is expressed by either a discrete value or a continuous value.
　この構成を備えることにより、本実施形態は、パラメータ１１１によって表現される統計モデルに対する尤度を計算することが可能となる。
　以下、これをより詳細に説明する。 With this configuration, the present embodiment can calculate the likelihood for the statistical model represented by the parameter 111.
Hereinafter, this will be described in more detail.
　図１は、本発明の第１の実施形態に係る尤度計算装置１００の構成を示す説明図である。尤度計算装置１００は一般的なコンピュータ装置であり、入力手段１０１、ＣＰＵ（Central Processing Unit）１０２、ＲＡＭ（Random Access Memory）１０３、ＨＤＤ（Hard Disk Drive）１０４、出力手段１０５を備える。入力手段１０１は一般的なキーボードなど、ユーザがデータ入力操作を行うことのできる装置である。ＣＰＵ１０２はコンピュータプログラムを実施する主体であり、後述する各プログラムがここで実行される。 FIG. 1 is an explanatory diagram showing a configuration of a likelihood calculating apparatus 100 according to the first embodiment of the present invention. The likelihood calculation device 100 is a general computer device, and includes an input unit 101, a CPU (Central Processing Unit) 102, a RAM (Random Access Memory) 103, an HDD (Hard Disk Drive) 104, and an output unit 105. The input unit 101 is a device that allows a user to perform data input operations, such as a general keyboard. The CPU 102 is a main body that executes a computer program, and each program described later is executed here.
　ＲＡＭ１０３は、ＣＰＵ１０２が実行する各プログラムおよび一時記憶データを記憶する揮発性記憶装置である。ＨＤＤ１０４は、各プログラムおよびデータを非実行時に収容する不揮発性記憶装置である。出力手段１０５は一般的なディスプレイなど、計算結果をユーザに示すことのできる装置である。 The RAM 103 is a volatile storage device that stores each program executed by the CPU 102 and temporary storage data. The HDD 104 is a non-volatile storage device that accommodates each program and data when not executed. The output means 105 is a device that can show a calculation result to the user, such as a general display.
　そして尤度計算手段１１０は、ＣＰＵ１０２がＨＤＤ１０４からＲＡＭ１０３にコンピュータプログラムを読み込むことによりソフトウェア上で構築されるものである。ユーザが入力手段１０１を通じてシンボルと遷移時間間隔とをもつデータ列を入力すると、尤度計算手段１１０はこのデータ列と統計モデルとを対応させ、統計モデルに対するデータ列の尤度を計算する。ここでいう尤度とは、データ列とモデルとの類似度を表す。ここで、この統計モデルはパラメータ１１１としてＨＤＤ１０４内に予め用意されている。 The likelihood calculating means 110 is constructed on software by the CPU 102 reading a computer program from the HDD 104 into the RAM 103. When the user inputs a data string having a symbol and a transition time interval through the input means 101, the likelihood calculating means 110 associates the data string with the statistical model and calculates the likelihood of the data string for the statistical model. The likelihood here represents the similarity between the data string and the model. Here, this statistical model is prepared in advance in the HDD 104 as the parameter 111.
　ユーザが入力するデータ列のデータ総個数をＮ、その中でｉ番目（１≦ｉ≦Ｎ）のデータを（Ｘｉ，Ｔｉ）とする。Ｘｉはｉ番目のシンボル、Ｔｉはｉ番目の遷移時間間隔を示す。Ｎ個のシンボルはＸ＝｛Ｘ１，Ｘ２，…Ｘｎ｝、Ｎ個の遷移時間間隔はＴ＝｛Ｔ１，Ｔ２，…Ｔｎ｝と表される。 Suppose that the total number of data in the data string input by the user is N, and the i-th (1 ≦ i ≦ N) data among them is (Xi, Ti). Xi represents the i-th symbol, and Ti represents the i-th transition time interval. N symbols are represented as X = {X1, X2,... Xn}, and N transition time intervals are represented as T = {T1, T2,.
　尤度計算手段１１０は、入力されたこれらの各データ列に対して、モデルに対する尤度Ｐ（Ｘ，Τ）を計算し、その計算結果を出力手段１０５へ出力する。 The likelihood calculation means 110 calculates the likelihood P (X, Τ) for the model for each of these input data strings, and outputs the calculation result to the output means 105.
　図２は、図１に示した尤度計算手段１１０の動作を表すフローチャートである。尤度計算手段１１０はまず、入力手段１０１からデータ列ＸおよびＴの入力を受け付ける（ステップＳ２０１）。続いて尤度計算手段１１０は、ＨＤＤ１０４からパラメータ１１１を読み込み（ステップＳ２０２）、後述の数３および数４によってデータ列ＸおよびＴの尤度を計算する（ステップＳ２０３）。そして、その計算結果を出力手段１０５へ出力する（ステップＳ２０４）。 FIG. 2 is a flowchart showing the operation of the likelihood calculating means 110 shown in FIG. The likelihood calculating means 110 first accepts input of data strings X and T from the input means 101 (step S201). Subsequently, the likelihood calculating unit 110 reads the parameter 111 from the HDD 104 (step S202), and calculates the likelihood of the data strings X and T using the following equations 3 and 4 (step S203). Then, the calculation result is output to the output means 105 (step S204).
　ここで、入力データ列をＸとし、データＸの長さをＮとする。入力データ列をＸ＝｛Ｘ１，Ｘ２，…ＸＮ｝とＴ＝｛τ１，τ２，…τＮ｝とする。ステップＳ２０３で行われるデータ列ＸおよびＴの尤度の計算は、次式で表される。
Figure JPOXMLDOC01-appb-M000003
 Here, the input data string is X, and the length of the data X is N. Assume that the input data strings are X = {X1, X2,... XN} and T = {τ1, τ2,. The likelihood calculation of the data strings X and T performed in step S203 is expressed by the following equation.
Figure JPOXMLDOC01-appb-M000003
　φｋは、コンポーネントk番目の遷移時間間隔の生起確率のパラメータを表す。Ｘはデータ列のシンボル列を表わす。データ列の長さをＮとする。Ｔは、データ列の遷移時間間隔の列を表わす。Ｋはデータ列ＸのＨＭＭの混合数であり、データ列Ｘが混合しない場合はＫ＝１となる。 Φk represents a parameter of the occurrence probability of the component k-th transition time interval. X represents a symbol string of the data string. Let N be the length of the data string. T represents a sequence of transition time intervals of the data sequence. K is the number of HMMs mixed in the data string X. When the data string X is not mixed, K = 1.
（第１の実施形態の全体的な動作）
　次に、上記の実施形態の全体的な動作について説明する。本発明に係る尤度計算方法は、シンボルと遷移時間間隔との組み合わせによって表現されるデータ系列に対して、データ系列の特定の統計モデルに対する尤度を計算する尤度計算方法であって、入力手段がデータ系列の入力を受け付け（図２：ステップＳ２０１）、記憶手段から統計モデルのパラメータを読み出し（図２：ステップＳ２０２）、パラメータを用いて尤度計算手段がデータ系列の統計モデルに対する尤度を計算し（図２：ステップＳ２０３）、計算された尤度を出力手段に出力する（図２：ステップＳ２０４）。 (Overall operation of the first embodiment)
Next, the overall operation of the above embodiment will be described. A likelihood calculation method according to the present invention is a likelihood calculation method for calculating a likelihood for a specific statistical model of a data sequence for a data sequence expressed by a combination of a symbol and a transition time interval, The means accepts the input of the data series (FIG. 2: step S201), reads the parameters of the statistical model from the storage means (FIG. 2: step S202), and the likelihood calculating means uses the parameters and the likelihood for the statistical model of the data series Is calculated (FIG. 2: step S203), and the calculated likelihood is output to the output means (FIG. 2: step S204).
　ここで、上記各動作ステップについては、これをコンピュータで実行可能にプログラム化し、これらを前記各ステップを直接実行するコンピュータである尤度計算装置１００に実行させるようにしてもよい。 Here, each of the above-described operation steps may be programmed to be executable by a computer, and may be executed by the likelihood calculating apparatus 100 which is a computer that directly executes each of the steps.
　この構成および動作により、本実施形態は以下のような効果を奏する。
　本実施形態によって、予めパラメータ１１１として与えられた統計モデルに対して、遷移時間間隔も考慮した尤度推定が可能となる。 With this configuration and operation, the present embodiment has the following effects.
According to this embodiment, it is possible to estimate the likelihood considering the transition time interval for the statistical model previously given as the parameter 111.
（第１の実施形態の変形例１）
　ここで、上述の第１の実施形態にはいくつかの変形が存在する。以下、それらの変形例について説明する。たとえば、ｎ個目の隠れ状態から遷移時間間隔が生起した確率を、時間間隔を離散化した値としてもよい。その場合は、下式で表される確率を多項分布としてよい。
Figure JPOXMLDOC01-appb-M000004
 (Modification 1 of the first embodiment)
Here, there are several variations in the first embodiment described above. Hereinafter, these modifications will be described. For example, the probability that the transition time interval has occurred from the nth hidden state may be a value obtained by discretizing the time interval. In that case, the probability represented by the following formula may be a multinomial distribution.
Figure JPOXMLDOC01-appb-M000004
（第１の実施形態の変形例２）
　また、数５で示したｋ番目のコンポーネントにおいてｎ個目の隠れ状態からｍ個目のデータ列のｎ個目の遷移時間間隔が生起した確率を、下式で示すように時間間隔は連続値としてもよい。この場合、下式の左辺で示される確率は連続分布を用いるとよい。下式は、指数分布とした場合である。
Figure JPOXMLDOC01-appb-M000005
 (Modification 2 of the first embodiment)
In addition, the probability that the nth transition time interval of the mth data string has occurred from the nth hidden state in the kth component shown in Equation 5 is a continuous value as shown in the following equation. It is good. In this case, the probability indicated by the left side of the following equation may be a continuous distribution. The following formula is for an exponential distribution.
Figure JPOXMLDOC01-appb-M000005
　上式では、連続値の分布として指数分布を用いた例を示したが、連続値の分布であればこれ以外の分布でも良い。ここで、αｉ（ｋ）は、コンポーネントｋ番目の指数分布のパラメータとする。その他の記号の意味は、第１の実施形態と同一である。 In the above equation, an example is shown in which an exponential distribution is used as the continuous value distribution, but other distributions may be used as long as the continuous value distribution is used. Here, αi (k) is a parameter of the component k-th exponential distribution. The meanings of the other symbols are the same as those in the first embodiment.
（第１の実施形態の変形例３）
　さらに、尤度計算手段１０３が数３に代えて下式を用いて計算を行うようにすることもできる。これは、時間間隔τｎが隠れ状態ＳｎおよびＳｎ－１から生起するとしたモデルである。
Figure JPOXMLDOC01-appb-M000006
 (Modification 3 of the first embodiment)
Further, the likelihood calculating means 103 may perform calculation using the following expression instead of the expression 3. This is a model in which the time interval τn arises from the hidden states Sn and Sn-1.
Figure JPOXMLDOC01-appb-M000006
　記号の意味は、第１の実施形態と同一である。τｎは、連続値であっても離散値であってもよい。 The meaning of the symbols is the same as in the first embodiment. τn may be a continuous value or a discrete value.
（第１の実施形態の変形例４）
　上述の変形例３で示した数６で、下式で示す要素を多項分布としてもよい。また、時間間隔を多項式としてもよい。この場合、下式の左辺で示される確率は連続分布を用いるとよい。下式は、指数分布とした場合である。
Figure JPOXMLDOC01-appb-M000007
 (Modification 4 of the first embodiment)
In the equation 6 shown in the above-described modification 3, the element shown by the following formula may be a multinomial distribution. The time interval may be a polynomial. In this case, the probability indicated by the left side of the following equation may be a continuous distribution. The following formula is for an exponential distribution.
Figure JPOXMLDOC01-appb-M000007
（第１の実施形態の変形例５）
　上述の第１の実施の形態では、隠れマルコフモデルにデータの時間間隔を組み入れたが、隠れ状態以外のマルコフモデルにデータの時間間隔を組み入れることもできる。下式に、マルコフモデルにデータの時間間隔を組み込むモデルを示す。
Figure JPOXMLDOC01-appb-M000008
 (Modification 5 of the first embodiment)
In the first embodiment described above, the data time interval is incorporated into the hidden Markov model. However, the data time interval may be incorporated into a Markov model other than the hidden state. The following equation shows a model that incorporates the data time interval into the Markov model.
Figure JPOXMLDOC01-appb-M000008
　ここで、Ｐ（χｎ）は、コンポーネントｋ番目のχｎが生起する確率を表す。Ｐ（χｎ｜χｎ－１，Ａｋ）は、コンポーネントｋ番目のχｎ－１からχｎが遷移する確率を表す。Ｐ（τｎ）（ｋ）は、コンポーネントｋ番目のτｎが生起する確率を表す。Ｐ（τｎ｜τｎ－１）は、コンポーネントｋ番目のτｎ－１からτｎが遷移する確率を表す。上記以外の記号の意味は、第１の実施形態と同一である。 Here, P (χn) represents the probability that the component k-th χn will occur. P (χn | χn−1, Ak) represents the probability that χn transitions from the component kth χn−1. P (τn) (k) represents the probability that the component k-th τn will occur. P (τn | τn−1) represents the probability of transition from τn−1 to τn of the component kth. The meanings of symbols other than the above are the same as those in the first embodiment.
（第２の実施形態）
　本発明の第２の実施形態に係る尤度計算装置３００は、第１の実施形態に係る尤度計算装置１００に、さらに入力されたデータ系列を用いて統計モデルのパラメータを推定して記憶手段に記憶するパラメータ推定手段３２０を追加したものである。
　この構成を備えることにより、本実施形態は、実際のデータから統計モデルのパラメータを学習することが可能となり、より的確に尤度を計算することが可能となる。
　以下、これをより詳細に説明する。 (Second Embodiment)
The likelihood calculating apparatus 300 according to the second embodiment of the present invention further estimates and stores the parameters of the statistical model using the data series input to the likelihood calculating apparatus 100 according to the first embodiment. The parameter estimation means 320 to be stored in is added.
With this configuration, the present embodiment can learn the parameters of the statistical model from actual data, and can more accurately calculate the likelihood.
Hereinafter, this will be described in more detail.
　図３は、本発明の第２の実施形態に係る尤度計算装置３００の構成を示す説明図である。尤度計算装置３００は、図１に示した第１の実施形態に係る尤度計算装置１００と同様のコンピュータ装置であり、尤度計算装置１００と同様の入力手段１０１、ＣＰＵ１０２、ＲＡＭ１０３、ＨＤＤ１０４、出力手段１０５を備える。同一の呼称および参照番号で呼んでいる動作部の機能および動作については、第１の実施形態として説明したものと同一である。 FIG. 3 is an explanatory diagram showing a configuration of a likelihood calculating apparatus 300 according to the second embodiment of the present invention. The likelihood calculation device 300 is a computer device similar to the likelihood calculation device 100 according to the first embodiment shown in FIG. 1, and has the same input means 101, CPU 102, RAM 103, HDD 104, and the like as the likelihood calculation device 100. Output means 105 is provided. The functions and operations of the operation units called with the same names and reference numbers are the same as those described as the first embodiment.
　本発明の第２の実施形態に係る尤度計算装置３００では、尤度計算手段１１０およびパラメータ推定手段３２０が、コンピュータプログラムをＣＰＵ１０２が実行することによりソフトウェア上で構築される。ユーザが、入力手段１０１を介して、第１の実施形態と同様のＮ個のシンボルＸ＝｛Ｘ１，Ｘ２，…Ｘｎ｝とＮ個の遷移時間間隔Ｔ＝｛Ｔ１，Ｔ２，…Ｔｎ｝とからなるデータ列を入力する。パラメータ推定手段３２０は、これらのデータ列からパラメータ１１１を推定し、これをＨＤＤ１０４に記憶させる。尤度計算手段１１０は、このパラメータ１１１を読み込んで、第１の実施形態で述べたと同一の動作でモデルに対するそのデータ列の尤度を計算する。 In the likelihood calculating apparatus 300 according to the second embodiment of the present invention, the likelihood calculating means 110 and the parameter estimating means 320 are constructed on software by the CPU 102 executing a computer program. The user inputs N symbols X = {X1, X2,... Xn} similar to the first embodiment and N transition time intervals T = {T1, T2,. Enter a data string consisting of. The parameter estimation unit 320 estimates the parameter 111 from these data strings and stores it in the HDD 104. The likelihood calculating means 110 reads the parameter 111 and calculates the likelihood of the data string for the model by the same operation as described in the first embodiment.
　図４は、図３に示したパラメータ推定手段３２０がパラメータ１１１を推定する動作について示すフローチャートである。パラメータ推定手段３２０はまず、入力手段１０１からデータ列ＸおよびＴの入力を受け付ける（ステップＳ４０１）。続いてパラメータ推定手段３２０は、前述の数１を用いてデータＸおよびＴをモデル化し、下式に示すＰを最大化するパラメータ１１１を公知のＥＭアルゴリズムを用いて推定する（ステップＳ４０２）。パラメータ推定手段３２０は、この計算で推定されたパラメータ１１１をＨＤＤ１０４に書き込んで記憶させる（ステップＳ４０３）。
Figure JPOXMLDOC01-appb-M000009
 FIG. 4 is a flowchart showing an operation in which the parameter estimation unit 320 shown in FIG. First, the parameter estimation unit 320 receives input of the data strings X and T from the input unit 101 (step S401). Subsequently, the parameter estimation unit 320 models the data X and T using the above-described equation 1, and estimates the parameter 111 that maximizes P shown in the following equation using a known EM algorithm (step S402). The parameter estimation unit 320 writes and stores the parameter 111 estimated by this calculation in the HDD 104 (step S403).
Figure JPOXMLDOC01-appb-M000009
　ここで、推定されるパラメータ１１１をθで表し、θ＝｛θ１，θ２，…，θＫ｝，θ（Ｋ）＝｛π（Ｋ），Γ（Ｋ），Ａ（Ｋ），Ｂ（Ｋ），φ（Ｋ）｝とする。その他の記号の意味は、数１および第１の実施の形態と同一である。尤度計算手段１１０は、このパラメータ１１１を読み込んで、第１の実施形態で述べたと同一の動作でモデルに対するそのデータ列の尤度を計算する。なお、この尤度の計算は、図４で示したパラメータ１１１を推定する動作と同時に行われてもよいし、別のタイミングで行われてもよい。 Here, the estimated parameter 111 is expressed by θ, θ = {θ1, θ2,..., ΘK}, θ (K) = {π (K), Γ (K), A (K), B (K). , Φ (K)}. The meanings of the other symbols are the same as those in Formula 1 and the first embodiment. The likelihood calculating means 110 reads the parameter 111 and calculates the likelihood of the data string for the model by the same operation as described in the first embodiment. This likelihood calculation may be performed simultaneously with the operation of estimating the parameter 111 shown in FIG. 4 or may be performed at another timing.
（第２の実施形態の全体的な動作）
　次に、上記の実施形態の全体的な動作について説明する。本発明に係る尤度計算方法は、第１の実施形態の動作に加えて、さらに入力されたデータ系列を用いてパラメータ推定手段が統計モデルのパラメータを推定し（図４：ステップＳ４０１～４０２）、推定されたパラメータをパラメータ推定手段が記憶手段に記憶する（図４：ステップＳ４０３）という動作が加わる。 (Overall operation of the second embodiment)
Next, the overall operation of the above embodiment will be described. In the likelihood calculation method according to the present invention, in addition to the operation of the first embodiment, the parameter estimation means estimates the parameters of the statistical model using the input data series (FIG. 4: steps S401 to 402). Then, an operation of storing the estimated parameters in the storage means by the parameter estimation means (FIG. 4: step S403) is added.
　ここで、上記各動作ステップについては、これをコンピュータで実行可能にプログラム化し、これらを前記各ステップを直接実行するコンピュータである尤度計算装置３００に実行させるようにしてもよい。 Here, each of the above-described operation steps may be programmed to be executable by a computer, and may be executed by the likelihood calculating apparatus 300 which is a computer that directly executes each of the steps.
　この構成および動作により、本実施形態は以下のような効果を奏する。
　本実施形態によって、実際のデータから統計モデルのパラメータ１１１を学習することが可能となり、より的確に尤度を計算することが可能となる。 With this configuration and operation, the present embodiment has the following effects.
According to this embodiment, it is possible to learn the parameter 111 of the statistical model from actual data, and it is possible to calculate the likelihood more accurately.
（第２の実施形態の変形例１）
　ここで、上述の第２の実施形態にはいくつかの変形が存在する。以下、それらの変形例について説明する。たとえば、前述のステップＳ４０２で用いる数１の、ｋ番目のコンポーネントにおいてｎ個目の隠れ状態からｍ個目のデータ列のｎ個目の遷移時間間隔が生起した確率を、時間間隔のシンボルを離散値として扱う数４に示した式によって求めてもよい。 (Modification 1 of 2nd Embodiment)
Here, there are some variations in the second embodiment described above. Hereinafter, these modifications will be described. For example, the probability that the n-th transition time interval of the m-th data string has occurred from the n-th hidden state in the k-th component of the equation 1 used in the above-described step S402 is the discrete time interval symbol. You may obtain | require by the type | formula shown to several 4 handled as a value.
（第２の実施形態の変形例２）
　また、ｋ番目のコンポーネントにおいてｎ個目の隠れ状態からｍ個目のデータ列のｎ個目の遷移時間間隔が生起した確率を、時間間隔のシンボルを連続値として扱う数５に示した式によって求めてもよい。 (Modification 2 of the second embodiment)
Further, the probability that the n-th transition time interval of the m-th data string has occurred in the k-th component from the n-th hidden state is expressed by the equation shown in Formula 5 in which symbols of the time interval are treated as continuous values. You may ask for it.
（第２の実施形態の変形例３）
　さらに、前述のステップＳ４０２で数１のかわりに数６に示した式を使用してもよい。 (Modification 3 of the second embodiment)
Further, the equation shown in Equation 6 may be used instead of Equation 1 in Step S402 described above.
（第２の実施形態の変形例４）
　また、この変形例３で、数６の要素である、ｍ個目のデータ列のｎ個目の遷移時間間隔τｍｎが、ｋ番目のコンポーネントのｎ－１個目の隠れ状態Ｓｎ－１およびｎ個目の隠れ状態Ｓｎから生起する確率を、数７に示した式で求めるようにすることもできる。 (Modification 4 of the second embodiment)
In the third modification, the n-th transition time interval τmn of the m-th data string, which is the element of Equation 6, is the n−1-th hidden state Sn−1 and n of the k-th component. The probability of occurrence from the first hidden state Sn can also be obtained by the equation shown in Equation 7.
（第３の実施形態）
　本発明の第３の実施形態は、第１および第２の実施形態に係る尤度計算装置（属性推定装置５００）で、入力手段１０１が、ユーザが訪問したウェブコンテンツの履歴を示すコンテンツ訪問履歴をシンボル、ユーザが一つのウェブコンテンツを訪問してから次のウェブコンテンツを訪問するまでのコンテンツ訪問間隔を遷移時間間隔として各々入力を受け付ける。そして尤度計算手段５１０が、計算された尤度をユーザの属性として出力する。 (Third embodiment)
The third embodiment of the present invention is a likelihood calculation device (attribute estimation device 500) according to the first and second embodiments, and the content visit history in which the input means 101 indicates the history of web content visited by the user. , And a content visit interval from the time when the user visits one web content to the next visit to the next web content as transition time intervals. Then, the likelihood calculating means 510 outputs the calculated likelihood as a user attribute.
　また、後述の変形例２にもあるように、このシンボルがユーザがウェブコンテンツを検索する際に使用した検索ワードの履歴を含むようにすることもできる。 Also, as in Modification 2 described later, this symbol can include a history of search words used when the user searches for web contents.
　この構成を備えることにより、本実施形態は、ウェブコンテンツの各ユーザの属性を尤度として算出することが可能となる。
　以下、これをより詳細に説明する。 With this configuration, the present embodiment can calculate the attribute of each user of web content as likelihood.
Hereinafter, this will be described in more detail.
　本発明の第３の実施形態は、以上で説明した第１および第２の実施形態に係る尤度計算装置を、ウェブサーバが各々のユーザに対して適したコンテンツ（ウェブページ）もしくは追加コンテンツ（広告や別のコンテンツへのリンクなど）を配信するために該ユーザの属性（嗜好、趣味、興味、年齢、性別、職業など）を推定する属性推定装置として使用するという実施形態である。 According to the third embodiment of the present invention, the likelihood calculation apparatus according to the first and second embodiments described above is a content (web page) or additional content (web page) suitable for each user by the web server. This is an embodiment in which it is used as an attribute estimation device that estimates the user's attributes (preference, hobby, interest, age, sex, occupation, etc.) in order to deliver advertisements or links to other contents.
　ウェブコンテンツを管理するシステムは、ユーザの属性に応じて、追加コンテンツとして配信する内容を選択する。例えば、追加コンテンツとして、以前にユーザが購入した商品と類似した商品の広告を配信する。多くの場合、この広告費から当該システムの運営費が捻出される。 The system for managing web content selects the content to be distributed as additional content according to the user attributes. For example, an advertisement of a product similar to a product previously purchased by the user is distributed as additional content. In many cases, the operating cost of the system is derived from this advertising cost.
　このため、ユーザにとって最適な追加コンテンツを配信するためには、ユーザの属性を適切に推定する必要がある。この推定が適切でないと、ユーザが全く興味のない追加コンテンツが配信されてしまう。例えば、スポーツに興味があるユーザには、スポーツ関連の商品広告を配信すると商品の購入までに至る可能性が高いが、そのようなユーザにアニメ関連の商品広告を配信しても商品の購入に至る可能性は低い。そればかりか、そのような自分の興味と異なる商品広告はユーザに不快感を与えてしまうので、当該ウェブコンテンツを利用しなくなってしまう可能性すらある。 For this reason, in order to deliver additional content optimal for the user, it is necessary to appropriately estimate the user's attributes. If this estimation is not appropriate, additional content that the user is not interested in will be distributed. For example, users who are interested in sports are more likely to purchase a product if they distribute a sports-related product advertisement. However, even if they distribute an animation-related product advertisement to such a user, they can purchase the product. Is unlikely. In addition, such product advertisements that are different from their own interests may make the user uncomfortable and may even prevent the web content from being used.
　そこで、本実施形態では、ユーザが過去にどのコンテンツをどの順番で訪問したのかというコンテンツ訪問履歴と、あるコンテンツに訪問してから次のコンテンツを訪問するまでのコンテンツ訪問間隔とを、入力されるデータ列とする。前者をシンボル、後者を遷移時間間隔とし、Ｎ個のシンボルＸ＝｛Ｘ１，Ｘ２，…Ｘｎ｝とＮ個の遷移時間間隔Ｔ＝｛Ｔ１，Ｔ２，…Ｔｎ｝とで表現されるデータ列であるものとする。 Therefore, in the present embodiment, the content visit history indicating which content the user has visited in the order and the content visit interval from visiting a certain content until visiting the next content is input. A data string. The former is a symbol, the latter is a transition time interval, and a data string represented by N symbols X = {X1, X2,... Xn} and N transition time intervals T = {T1, T2,. It shall be.
　図５は、本発明の第３の実施形態に係る属性推定装置５００の構成を示す説明図である。属性推定装置５００は、図３に示した第２の実施形態に係る尤度計算装置３００と同様のコンピュータ装置であり、尤度計算装置３００と同様の入力手段１０１、ＣＰＵ１０２、ＲＡＭ１０３、ＨＤＤ１０４、出力手段１０５を備える。同一の呼称および参照番号で呼んでいる動作部の機能および動作については、第１および第２の実施形態として説明したものと同一である。 FIG. 5 is an explanatory diagram showing a configuration of an attribute estimation apparatus 500 according to the third embodiment of the present invention. The attribute estimation apparatus 500 is a computer apparatus similar to the likelihood calculation apparatus 300 according to the second embodiment shown in FIG. 3, and has the same input means 101, CPU 102, RAM 103, HDD 104, output as the likelihood calculation apparatus 300. Means 105 is provided. The functions and operations of the operation units called with the same names and reference numbers are the same as those described as the first and second embodiments.
　属性推定装置５００では、尤度計算手段５１０および属性モデル学習手段５２０が、コンピュータプログラムをＣＰＵ１０２が実行することによりソフトウェア上で構築される。入力手段１０１からこれらの動作手段に対して、前述のように、第１～２の実施形態と同様に、Ｍ個のシンボルとＭ個の遷移時間間隔とからなるデータ列が入力される。 In the attribute estimation apparatus 500, the likelihood calculation means 510 and the attribute model learning means 520 are constructed on software by the CPU 102 executing a computer program. As described above, a data string composed of M symbols and M transition time intervals is input from the input unit 101 to these operation units as in the first and second embodiments.
　尤度計算手段５１０および属性モデル学習手段５２０は各々、第２の実施形態に係る尤度計算装置３００における尤度計算手段１１０およびパラメータ推定手段３２０と同様に動作する。そして、学習したモデルのパラメータ１１１をＨＤＤ１０４に書き込む。各属性モデルには、第１の実施形態で説明した数３を用いる。そして、属性ごとのパラメータ１１１は、第２の実施形態の図４のステップ４０２で説明したθおよびθ（Ｋ）である。 Likelihood calculation means 510 and attribute model learning means 520 operate in the same manner as likelihood calculation means 110 and parameter estimation means 320 in likelihood calculation apparatus 300 according to the second embodiment. Then, the learned model parameter 111 is written in the HDD 104. For each attribute model, Equation 3 described in the first embodiment is used. And the parameter 111 for every attribute is (theta) and (theta) (K) demonstrated in step 402 of FIG. 4 of 2nd Embodiment.
　図６は、図５で示した属性モデル学習手段５２０が、属性ごとに学習したモデルのパラメータの例を示す説明図である。このように、推定候補となる各属性ごとのモデルを表すパラメータθを学習する処理を、属性モデル学習処理と呼ぶ。 FIG. 6 is an explanatory diagram showing an example of model parameters learned for each attribute by the attribute model learning means 520 shown in FIG. In this manner, the process of learning the parameter θ representing the model for each attribute that is a candidate for estimation is referred to as an attribute model learning process.
　尤度計算手段５１０は、ユーザのコンテンツ訪問履歴Ｘとコンテンツ訪問間隔Ｔが入力されると、属性を推定するために必要なパラメータ１１１を推定候補となる属性分だけＨＤＤ１０４から読み出し、ユーザの属性を推定する。そして、推定したユーザの属性を候補の中から決定して、その結果を出力手段１０５に出力する。 When the user's content visit history X and the content visit interval T are input, the likelihood calculating means 510 reads the parameter 111 necessary for estimating the attribute from the HDD 104 by the attribute that is the estimation candidate, and extracts the user's attribute. presume. Then, the estimated user attribute is determined from the candidates, and the result is output to the output means 105.
　図７は、図５で示した尤度計算手段５１０および属性モデル学習手段５２０に入力手段１０１から入力される、コンテンツ訪問履歴Ｘとコンテンツ訪問間隔Ｔの例を示す説明図である。図７では、コンテンツ訪問間隔Ｔを、ＭＩＤ、ＬＯＷなどのような離散シンボルとして扱う。例えば、訪問間隔が１分以内の場合はＬＯＷ、１時間以内の場合はＭＩＤというように、予め定義されている。 FIG. 7 is an explanatory diagram showing an example of the content visit history X and the content visit interval T input from the input unit 101 to the likelihood calculating unit 510 and the attribute model learning unit 520 shown in FIG. In FIG. 7, the content visit interval T is treated as a discrete symbol such as MID or LOW. For example, when the visit interval is within 1 minute, LOW is defined, and when the visit interval is within 1 hour, MID is defined in advance.
　図７のコンテンツ訪問履歴を示す矢印（→）は、あるコンテンツから次のコンテンツを訪問する遷移を表す。さらに、コンテンツ訪問間隔を示す矢印（→）は、あるコンテンツから次のコンテンツを訪問するまでの訪問間隔の遷移を表す。ＨＤＤ１０４に記憶されるパラメータ１１１は、属性モデル学習手段５２０によって学習される各属性モデルのパラメータθである。 The arrow (→) indicating the content visit history in FIG. 7 represents a transition from one content to the next content. Furthermore, an arrow (→) indicating a content visit interval represents a transition of a visit interval from a certain content until a next content is visited. The parameter 111 stored in the HDD 104 is the parameter θ of each attribute model learned by the attribute model learning means 520.
　尤度計算手段５１０によって推定されたユーザの属性は、最終的に出力手段１０５に出力される。 The user attributes estimated by the likelihood calculation means 510 are finally output to the output means 105.
　図８は、図５で示した属性モデル学習手段５２０が行う属性モデル学習処理を表すフローチャートである。ユーザは、入力手段１０１を介して、第１～２の実施形態と同様にして、予め属性が判明しているユーザのＮ個のコンテンツ訪問履歴Ｘ＝｛Ｘ１，Ｘ２，…Ｘｎ｝とＮ個のコンテンツ訪問間隔Ｔ＝｛Ｔ１，Ｔ２，…Ｔｎ｝とからなるデータ列を入力する。属性モデル学習手段５２０はこのデータ列ＸおよびＴの入力を受け付ける（ステップＳ８０１）。 FIG. 8 is a flowchart showing the attribute model learning process performed by the attribute model learning means 520 shown in FIG. As in the first and second embodiments, the user can use N content visit histories X = {X 1, X 2,. A data string consisting of the content visit interval T = {T1, T2,... Tn} is input. The attribute model learning unit 520 receives the input of the data strings X and T (step S801).
　続いて属性モデル学習手段５２０は、入力されたユーザの属性に対応した属性モデルのパラメータ１１１をＨＤＤ１０４から読み出す（ステップＳ８０２）。入力したユーザの属性に対応した属性モデルのパラメータを一度も推定していない場合は、初期値のパラメータを読み出すこととする。 Subsequently, the attribute model learning unit 520 reads the attribute model parameter 111 corresponding to the input user attribute from the HDD 104 (step S802). When the parameter of the attribute model corresponding to the input user attribute has never been estimated, the initial value parameter is read out.
　続いて属性モデル学習手段５２０は、入力されたユーザのコンテンツ訪問履歴Ｘとコンテンツ時間間隔Ｔと、パラメータ１１１とを用いて、属性モデルを学習する（ステップＳ８０３）。ここでいう属性モデルの学習とは、図４のステップＳ４０２で述べた、パラメータを推定することを同義である。第１の実施形態で説明した数３を各属性モデルの属性モデルとして使用し、数９で示したＰを最大化するパラメータ１１１を公知のＥＭアルゴリズムを用いて推定する。最後に、属性モデル学習手段５２０は、この計算で推定されたパラメータ１１１をＨＤＤ１０４に書き込んで記憶させる（ステップＳ８０４）。
次に、属性推定処理について説明する。 Subsequently, the attribute model learning unit 520 learns the attribute model using the input content visit history X of the user, the content time interval T, and the parameter 111 (step S803). The learning of the attribute model here is synonymous with the parameter estimation described in step S402 in FIG. The equation 3 described in the first embodiment is used as the attribute model of each attribute model, and the parameter 111 that maximizes P shown in the equation 9 is estimated using a known EM algorithm. Finally, the attribute model learning unit 520 writes and stores the parameter 111 estimated by this calculation in the HDD 104 (step S804).
Next, attribute estimation processing will be described.
　図９は、図５で示した尤度計算手段５１０が行うユーザ属性の推定処理を表すフローチャートである。尤度計算手段５１０は、図８で属性モデル学習手段５２０に入力されたものと同一のこのデータ列ＸおよびＴの入力を受け付ける（ステップＳ９０１）。このデータ列に対して、尤度計算手段５１０は属性推定の候補となる属性モデルのパラメータ１１１を読み出し（ステップＳ９０２）、各属性モデルに対する尤度Ｐ（Ｘ，Ｔ）を算出する（ステップＳ９０３）。この尤度の計算は、第１の実施形態の尤度計算手段１０３が、図２のステップＳ２０３で行った計算の内容と同一である。 FIG. 9 is a flowchart showing a user attribute estimation process performed by the likelihood calculating means 510 shown in FIG. The likelihood calculating means 510 receives the same data strings X and T as those inputted to the attribute model learning means 520 in FIG. 8 (step S901). For this data string, the likelihood calculation means 510 reads the parameter 111 of the attribute model that is a candidate for attribute estimation (step S902), and calculates the likelihood P (X, T) for each attribute model (step S903). . This likelihood calculation is the same as the content of the calculation performed by the likelihood calculating means 103 of the first embodiment in step S203 of FIG.
　図１０は、図９のステップＳ９０３で算出された、各属性に対する尤度の例を示す説明図である。例えば、「スポーツ」、「アニメ」、「ニュース」、「旅行」などのようにユーザの嗜好を表すジャンル名を属性の候補とすると、入力された推定するユーザのコンテンツ訪問履歴とコンテンツ訪問間隔に対する尤度が、図１０のように出力される。 FIG. 10 is an explanatory diagram showing an example of likelihood for each attribute calculated in step S903 in FIG. For example, if genre names representing user preferences such as “sports”, “animation”, “news”, “travel”, and the like are attribute candidates, the estimated content visit history and the content visit interval of the user are input. The likelihood is output as shown in FIG.
　図９に戻って、尤度計算手段５１０は、ステップＳ９０３で算出された尤度から、ユーザの属性を推定する（ステップＳ９０４）。たとえば最も高い尤度を出力した属性モデルの属性をそのユーザの推定結果とすることができる。図１０に示した例では、「スポーツ」という属性の尤度が最も高いため、推定結果は、「スポーツ」となる。そして尤度計算手段５１０は、推定されたユーザの属性を、出力手段１０５に出力する（ステップＳ９０５）。 Returning to FIG. 9, the likelihood calculating means 510 estimates the user attribute from the likelihood calculated in step S903 (step S904). For example, the attribute of the attribute model that outputs the highest likelihood can be used as the estimation result of the user. In the example illustrated in FIG. 10, since the likelihood of the attribute “sport” is the highest, the estimation result is “sport”. The likelihood calculating unit 510 outputs the estimated user attribute to the output unit 105 (step S905).
　図１１は、図９のステップＳ９０５で出力された、ユーザの属性の出力の例を示す説明図である。図１１に示した例では、このユーザの属性が「スポーツ」であるという旨の出力がなされている。
（第３の実施形態の全体的な動作）
　次に、上記の実施形態の全体的な動作について説明する。本実施形態に係る動作は、第１～第２の実施形態の動作で、尤度計算手段１１０が、計算された尤度をユーザの属性として出力する。 FIG. 11 is an explanatory diagram showing an example of the output of the user attribute output in step S905 of FIG. In the example illustrated in FIG. 11, an output indicating that the user attribute is “sports” is output.
(Overall operation of the third embodiment)
Next, the overall operation of the above embodiment will be described. The operation according to the present embodiment is the operation of the first and second embodiments, and the likelihood calculating means 110 outputs the calculated likelihood as a user attribute.
　ここで、上記各動作ステップについては、これをコンピュータで実行可能にプログラム化し、これらを前記各ステップを直接実行するコンピュータである属性推定装置５００に実行させるようにしてもよい。 Here, the above-described operation steps may be programmed to be executable by a computer, and may be executed by the attribute estimation apparatus 500 which is a computer that directly executes the respective steps.
　この構成および動作により、本実施形態は以下のような効果を奏する。
　本実施形態によって、ユーザの嗜好、趣味、興味、年齢、性別、職業などといった属性を的確に推定して、コンテンツの配信に利用することができる。 With this configuration and operation, the present embodiment has the following effects.
According to the present embodiment, attributes such as user preferences, hobbies, interests, age, sex, occupation, etc. can be accurately estimated and used for content distribution.
（第３の実施形態の変形例１）
　ここで、上述の第３の実施形態にはいくつかの変形が存在する。以下、それらの変形例について説明する。たとえば、入力手段１０１から入力されるコンテンツ訪問時間間隔Ｔを、離散シンボルではなく、連続値としてもよい。図１２は、本変形例で入力手段１０１から入力される、コンテンツ訪問履歴Ｘとコンテンツ訪問間隔Ｔの例を示す説明図である。ここではコンテンツ訪問時間間隔Ｔが、ＭＩＤ、ＬＯＷなどのような離散シンボルではなく、具体的な秒数で示されている。 (Modification 1 of 3rd Embodiment)
Here, there are some variations in the above-described third embodiment. Hereinafter, these modifications will be described. For example, the content visit time interval T input from the input unit 101 may be a continuous value instead of a discrete symbol. FIG. 12 is an explanatory diagram illustrating an example of the content visit history X and the content visit interval T input from the input unit 101 in the present modification. Here, the content visit time interval T is not a discrete symbol such as MID or LOW but a specific number of seconds.
（第３の実施形態の変形例２）
　また、コンテンツ訪問履歴Ｘとコンテンツ訪問間隔Ｔに加えて、検索ワード履歴をユーザの属性推定に使用することもできる。ここでいう検索ワードとは、ユーザが当該コンテンツを検索する際に、Ｙａｈｏｏ（登録商標）、Ｇｏｏｇｌｅ（登録商標）などのような検索エンジンに入力した検索キーワードとなる単語である。図１３は、本変形例で入力手段１０１から入力される、コンテンツ訪問履歴Ｘとコンテンツ訪問間隔Ｔおよび検索ワードの例を示す説明図である。 (Modification 2 of the third embodiment)
In addition to the content visit history X and the content visit interval T, the search word history can also be used for user attribute estimation. The search word here is a word that becomes a search keyword input to a search engine such as Yahoo (registered trademark) or Google (registered trademark) when the user searches the content. FIG. 13 is an explanatory diagram illustrating examples of the content visit history X, the content visit interval T, and the search word that are input from the input unit 101 in the present modification.
（第３の実施形態の変形例３）
　さらに、予めパラメータ１１１が学習されてＨＤＤ１０４に記憶されていれば、属性推定装置５００は、属性モデル学習手段５２０を省略した構成とすることもできる。 (Modification 3 of the third embodiment)
Furthermore, if the parameter 111 is learned in advance and stored in the HDD 104, the attribute estimation device 500 may be configured such that the attribute model learning unit 520 is omitted.
（第４の実施形態）
　本発明の第４の実施形態は、以上で説明した第３の実施形態に係る属性推定装置５００を含むコンテンツ配信システムであり、そこで推定されたユーザの属性に応じて、ウェブサーバがそのユーザの属性に対応した広告情報をそのコンテンツに付け加えて配信するという実施形態である。 (Fourth embodiment)
The fourth embodiment of the present invention is a content distribution system that includes the attribute estimation apparatus 500 according to the third embodiment described above, and the web server determines the user's attribute according to the user attribute estimated there. In this embodiment, advertisement information corresponding to an attribute is added to the content and distributed.
　より具体的には、本発明の第４の実施形態に係るコンテンツ配信システム１４００は、ユーザが操作するユーザ端末１４０１と、ウェブサーバ１４０３と、第３の実施形態に係る属性推定装置５００とがネットワーク（インターネット１４０２）を介して相互に接続されたコンテンツ配信システムである。ウェブサーバ１４０３は、ユーザ端末１４０１からのリクエストを受け付けるリクエスト受付手段１６１２と、リクエストに対応したコンテンツをユーザ端末に配信するコンテンツ配信手段１６１４と、ユーザにリクエストされたコンテンツの履歴を示すコンテンツ訪問履歴およびユーザが一つのコンテンツから次のコンテンツを訪問するまでのコンテンツ訪問間隔とを記憶するユーザリクエスト記憶部１６２３と、コンテンツ訪問履歴をシンボル、コンテンツ訪問間隔を遷移時間間隔とするデータ系列として尤度計算装置５００に入力する出力手段１６０６とを有し、さらに尤度計算装置で計算されたユーザの属性の入力を受け付ける入力手段１６０１と、ユーザの属性に基づいてコンテンツに追加する広告を選択する広告選択手段１６１１とを有し、コンテンツ配信手段１６１４がコンテンツに広告を追加してユーザ端末に配信する。 More specifically, a content distribution system 1400 according to the fourth embodiment of the present invention includes a user terminal 1401 operated by a user, a web server 1403, and an attribute estimation apparatus 500 according to the third embodiment. This is a content distribution system connected to each other via (Internet 1402). The web server 1403 includes a request reception unit 1612 that receives a request from the user terminal 1401, a content distribution unit 1614 that distributes content corresponding to the request to the user terminal, a content visit history that indicates a history of the content requested by the user, and A user request storage unit 1623 for storing a content visit interval until a user visits the next content from one content, and a likelihood calculation device as a data series having a content visit history as a symbol and a content visit interval as a transition time interval 500, an input unit 1601 that receives an input of a user attribute calculated by the likelihood calculating device, and an advertisement selection unit that selects an advertisement to be added to the content based on the user attribute 161 Has a door, content delivery means 1614 is delivered to the user terminal by adding advertising to the content.
　この構成を備えることにより、本実施形態は、算出された各ユーザの属性に適する内容の広告をユーザ端末に配信することが可能となる。
　以下、これをより詳細に説明する。 By providing this configuration, the present embodiment can distribute advertisements with contents suitable for the calculated attributes of each user to the user terminals.
Hereinafter, this will be described in more detail.
　本発明の第４の実施形態は、ユーザがコンテンツ（ウェブページ）をサーバにリクエストした際に、ユーザの属性に対応した広告情報をそのコンテンツに付け加えたものを、サーバからユーザに配信する。その際、ユーザの属性の推定に、第３の実施の形態に係る属性推定装置５００を利用する。 In the fourth embodiment of the present invention, when a user requests content (web page) from a server, advertisement information corresponding to the user attribute is added to the content and distributed from the server to the user. In that case, the attribute estimation apparatus 500 which concerns on 3rd Embodiment is utilized for estimation of a user's attribute.
　図１４は、本発明の第４の実施形態に係るコンテンツ配信システム１４００で、各装置の接続構成を示す説明図である。このコンテンツ配信システムは、ユーザに対して入出力等のインターフェイスを提供するコンピュータ装置であるユーザ端末１４０１と、ユーザからリクエストされたコンテンツにユーザの属性に対応した広告情報をそのコンテンツに付け加えたものをユーザに配信するコンピュータ装置であるウェブサーバ１４０３と、第３の実施の形態に係る属性推定装置５００とがインターネット１４０２で相互に接続されて構成されている。 FIG. 14 is an explanatory diagram showing a connection configuration of each device in the content distribution system 1400 according to the fourth embodiment of the present invention. This content distribution system includes a user terminal 1401 which is a computer device that provides an input / output interface to a user, and content obtained by adding advertisement information corresponding to a user attribute to the content requested by the user. A web server 1403 that is a computer device distributed to a user and an attribute estimation device 500 according to the third embodiment are configured to be connected to each other via the Internet 1402.
　図１５は、図１４で示したユーザ端末１４０１の構成を示す説明図である。ユーザ端末１４０１は一般的なパーソナルコンピュータなどのようなコンピュータ装置であり、入力手段１５０１、ＣＰＵ１５０２、ＲＡＭ１５０３、ネットワークカード１５０４、出力手段１５０５を備える。ネットワークカード１５０４は、インターネット１４０２に接続して通信を行うインターフェイスである。これら以外の各構成部の機能および動作は、図１、図３、図５で説明した尤度計算装置および属性推定装置における同名の機能部と同一である。 FIG. 15 is an explanatory diagram showing the configuration of the user terminal 1401 shown in FIG. The user terminal 1401 is a computer device such as a general personal computer, and includes an input unit 1501, a CPU 1502, a RAM 1503, a network card 1504, and an output unit 1505. The network card 1504 is an interface that communicates by connecting to the Internet 1402. The functions and operations of the respective components other than these are the same as the functional units having the same names in the likelihood calculation device and the attribute estimation device described with reference to FIGS. 1, 3, and 5.
　ユーザ端末１４０１では、コンテンツリクエスト手段１５１１、コンテンツ表示手段１５１２が、コンピュータプログラムをＣＰＵ１５０２が実行することによりソフトウェア上で構築される。コンテンツリクエスト手段１５１１は、ユーザが訪問したいコンテンツについての入力を入力手段１５０１を通じて受け付け、このリクエストをネットワークカード１５０４を介してウェブサーバ１４０３にリクエストする。例えば、スポーツのコンテンツを訪問したい場合は、ウェブサーバにスポーツのコンテンツをリクエストする。 In the user terminal 1401, the content request unit 1511 and the content display unit 1512 are constructed on software by the CPU 1502 executing a computer program. The content request unit 1511 receives an input about the content that the user wants to visit through the input unit 1501, and requests this request to the web server 1403 through the network card 1504. For example, when a user wants to visit sports content, he / she requests the sports content from the web server.
　コンテンツ表示手段１５１２は、ウェブサーバ１４０３から当該リクエストに応じて返送されてきたコンテンツを、ネットワークカード１５０４を介して受信し、出力手段１５０５に表示する。 The content display unit 1512 receives the content returned from the web server 1403 in response to the request via the network card 1504 and displays it on the output unit 1505.
　図１６は、図１４で示したウェブサーバ１４０３の構成を示す説明図である。ウェブサーバ１４０３は一般的なサーバコンピュータなどのようなコンピュータ装置であり、入力手段１６０１、ＣＰＵ１６０２、ＲＡＭ１６０３、ＨＤＤ１６０４、ネットワークカード１６０５、出力手段１６０６を備える。ウェブサーバ１４０３は、ユーザ端末１４０１からリクエストされたコンテンツに、ユーザの属性に対応した広告情報を付け加えてユーザ端末１４０１に配信するという機能を実行する。 FIG. 16 is an explanatory diagram showing the configuration of the web server 1403 shown in FIG. The web server 1403 is a computer device such as a general server computer, and includes an input unit 1601, a CPU 1602, a RAM 1603, an HDD 1604, a network card 1605, and an output unit 1606. The web server 1403 executes a function of adding advertisement information corresponding to the user attribute to the content requested from the user terminal 1401 and delivering the advertisement information to the user terminal 1401.
　ネットワークカード１５０４は、図１５のユーザ端末１４０１と同じく、インターネット１４０２に接続して通信を行うインターフェイスである。入力手段１６０１および出力手段１６０６は、属性推定装置５００への情報の入出力を行う。入力手段１６０１および出力手段１６０６の属性推定装置５００への接続方法は、任意の接続方法が適用できる。これら以外の各構成部の機能および動作は、図１、図３、図５で説明した尤度計算装置および属性推定装置における同名の機能部と同一である。 The network card 1504 is an interface that communicates by connecting to the Internet 1402, similar to the user terminal 1401 of FIG. 15. The input unit 1601 and the output unit 1606 input / output information to / from the attribute estimation apparatus 500. As a method for connecting the input unit 1601 and the output unit 1606 to the attribute estimation apparatus 500, any connection method can be applied. The functions and operations of the respective components other than these are the same as the functional units having the same names in the likelihood calculation device and the attribute estimation device described with reference to FIGS. 1, 3, and 5.
　ウェブサーバ１４０３では、広告選択手段１６１１、リクエスト受付手段１６１２、読み出し手段１６１３、コンテンツ配信手段１６１４が、コンピュータプログラムをＣＰＵ１６０２が実行することによりソフトウェア上で構築される。そしてＨＤＤ１６０４には、広告記憶部１６２１、コンテンツ記憶部１６２２、ユーザリクエスト記憶部１６２３といった各情報の記憶部が存在する。以下、その各々の内容について説明する。 In the web server 1403, the advertisement selection unit 1611, the request reception unit 1612, the reading unit 1613, and the content distribution unit 1614 are constructed on software by the CPU 1602 executing a computer program. The HDD 1604 includes storage units for information such as an advertisement storage unit 1621, a content storage unit 1622, and a user request storage unit 1623. Hereinafter, each content will be described.
　広告選択手段１６１１は、ユーザの属性推定結果をユーザ属性推定装置５００から受け取り、その推定結果に対応したユーザの属性に対応する広告を広告記憶部１６２１から取り出す。そして、コンテンツ配信手段１６１４からユーザに配信するコンテンツに、広告記憶部１６２１から取り出した広告を付け加える。 The advertisement selection unit 1611 receives the user attribute estimation result from the user attribute estimation apparatus 500 and extracts the advertisement corresponding to the user attribute corresponding to the estimation result from the advertisement storage unit 1621. Then, the advertisement extracted from the advertisement storage unit 1621 is added to the content distributed from the content distribution unit 1614 to the user.
　リクエスト受付手段１６１２は、ユーザ端末１４０１からコンテンツのリクエストを受け付ける。読み出し手段１６１３は、リクエストをしたユーザの過去のコンテンツ訪問履歴とコンテンツ訪問間隔をユーザリクエスト記憶部１６２３から読み出す。 The request accepting unit 1612 accepts a content request from the user terminal 1401. The reading unit 1613 reads from the user request storage unit 1623 the past content visit history and content visit interval of the user who made the request.
　コンテンツ配信手段１６１４は、ユーザ端末１４０１からリクエストされたコンテンツにユーザの属性に対応した広告情報を付け加えたものを、ユーザ端末１４０１に配信する。 The content distribution unit 1614 distributes the content requested from the user terminal 1401 to which the advertisement information corresponding to the user attribute is added to the user terminal 1401.
　ユーザリクエスト記憶部１６２３は、ユーザからリクエストされた内容を、ユーザ名、リクエストされた時刻とともに記憶する。コンテンツ記憶部１６２２は、コンテンツ配信手段によってユーザに配信するコンテンツを記憶する。広告記憶部１６２１は、広告選択手段１６１１が配信する広告を記憶する。 The user request storage unit 1623 stores the content requested by the user together with the user name and the requested time. The content storage unit 1622 stores the content distributed to the user by the content distribution unit. The advertisement storage unit 1621 stores advertisements distributed by the advertisement selection unit 1611.
　入力手段１６０１は、属性推定装置５００から入力されたユーザ属性推定結果を広告選択手段１６１１に渡す。出力手段１６０６は、読み出し手段１６１３によって取り出されたユーザの過去のコンテンツ訪問履歴とコンテンツ訪問間隔を属性推定装置５００に出力する。 The input unit 1601 passes the user attribute estimation result input from the attribute estimation device 500 to the advertisement selection unit 1611. The output unit 1606 outputs the user's past content visit history and content visit interval extracted by the reading unit 1613 to the attribute estimation apparatus 500.
　属性推定装置５００は、図５に示したものと同様の構成であり、第３の実施の形態として既に説明した動作を行う。 The attribute estimation apparatus 500 has the same configuration as that shown in FIG. 5 and performs the operation already described as the third embodiment.
　次に、第４の実施形態に係る動作について説明する。動作は、属性モデル学習処理と推定した属性に対応する広告の配信処理の２つに分かれる。まず、属性モデル学習処理について説明する。図１７は、図１４～１６に示したユーザ端末１４０１、ウェブサーバ１４０３、および属性推定装置５００が行う属性モデル学習処理の動作を示すシーケンス図である。 Next, the operation according to the fourth embodiment will be described. The operation is divided into an attribute model learning process and an advertisement distribution process corresponding to the estimated attribute. First, the attribute model learning process will be described. FIG. 17 is a sequence diagram showing the operation of the attribute model learning process performed by the user terminal 1401, the web server 1403, and the attribute estimation apparatus 500 shown in FIGS.
　まず、予め属性が判明しているユーザが、ユーザ端末１４０１のコンテンツリクエスト手段１５１１を操作し、ウェブサーバ１４０３にコンテンツをリクエストする（ステップＳ１７０１）。ウェブサーバ１４０３では、リクエスト受付手段１６１２がこのリクエストを受け付けてユーザリクエスト記憶部１６２３に記憶する（ステップＳ１７０２）。 First, a user whose attribute is known in advance operates the content request unit 1511 of the user terminal 1401, and requests content from the web server 1403 (step S1701). In the web server 1403, the request reception unit 1612 receives this request and stores it in the user request storage unit 1623 (step S1702).
　そして、コンテンツ配信手段１６１４がリクエストされたコンテンツをコンテンツ記憶部１６２２から読み出し、コンテンツ配信手段１６１４がこれをユーザ端末１４０１に配信する（ステップＳ１７０３）。ユーザ端末１４０１では、コンテンツ表示手段１５１２が配信されたこのコンテンツを受け取って表示する（ステップＳ１７０４）。 Then, the content distribution unit 1614 reads the requested content from the content storage unit 1622, and the content distribution unit 1614 distributes it to the user terminal 1401 (step S1703). In the user terminal 1401, the content display unit 1512 receives and displays the distributed content (step S1704).
　そしてコンテンツ配信手段１６１４は、リクエストを受け付けたユーザの情報を読み出し手段１６１３に出力する（ステップＳ１７０５）。ここでいう「リクエストを受け付けたユーザの情報」とは、ユーザリクエスト記憶部１６２３からそのユーザの過去のコンテンツ訪問履歴とコンテンツ訪問間隔を読み出せる、ユーザと履歴とが一対一に対応しうる情報であればよい。 Then, the content distribution unit 1614 outputs the information of the user who accepted the request to the reading unit 1613 (step S1705). The “information of the user who accepted the request” here is information that can read the user's past content visit history and content visit interval from the user request storage unit 1623, and the user and the history can correspond one-to-one. I just need it.
　そして読み出し手段１６１３は、ユーザリクエスト記憶部１６２３から、リクエストを受け付けたユーザの過去のコンテンツ訪問履歴とコンテンツ訪問間隔を読み出す（ステップＳ１７０６）。さらに読み出し手段１６１３は、読み出したユーザの過去のコンテンツ訪問履歴とコンテンツ訪問間隔、予め判明しているユーザ属性を、出力手段１６０６を通じて属性推定装置５００に入力する（ステップＳ１７０７）。 Then, the reading unit 1613 reads from the user request storage unit 1623 the past content visit history and content visit interval of the user who accepted the request (step S1706). Furthermore, the reading unit 1613 inputs the read content visit history of the user, the content visit interval, and the previously determined user attribute to the attribute estimation apparatus 500 through the output unit 1606 (step S1707).
　属性推定装置５００は、入力されたユーザの属性に対応した属性モデルのパラメータ１１１をＨＤＤ１０４から読み出す。入力したユーザの属性に対応した属性モデルを一度も学習していない場合は、初期値のパラメータ１１１をＨＤＤ１０４から読み出すこととする。 The attribute estimation apparatus 500 reads the parameter 111 of the attribute model corresponding to the input user attribute from the HDD 104. If the attribute model corresponding to the input user attribute has never been learned, the initial value parameter 111 is read from the HDD 104.
　属性推定装置５００では、属性モデル学習手段５２０が、入力されたユーザの過去のコンテンツ訪問履歴とコンテンツ時間間隔とパラメータ１１１とを用いて、属性モデルを学習する（ステップＳ１７０８～１７０９）。そして、学習結果に基づくパラメータ１１１をＨＤＤ１０４に書き込む（ステップＳ１７１０）。以上で、属性モデル学習処理は完了する。 In the attribute estimation apparatus 500, the attribute model learning means 520 learns an attribute model using the input past content visit history, content time interval, and parameter 111 of the user (steps S1708 to 1709). Then, the parameter 111 based on the learning result is written in the HDD 104 (step S1710). This completes the attribute model learning process.
　次に、推定した属性に対応する広告の配信処理について説明する。図１８は、図１４～１６に示したユーザ端末１４０１、ウェブサーバ１４０３、および属性推定装置５００が行う、推定した属性に対応する広告の配信処理の動作の例を示すシーケンス図である。 Next, an advertisement distribution process corresponding to the estimated attribute will be described. FIG. 18 is a sequence diagram illustrating an example of the operation of the advertisement distribution process corresponding to the estimated attribute performed by the user terminal 1401, the web server 1403, and the attribute estimation apparatus 500 illustrated in FIGS.
　まず、ユーザがユーザ端末１４０１のコンテンツリクエスト手段１５１１を操作し、ウェブサーバ１４０３にコンテンツをリクエストする（ステップＳ１８０１）。ウェブサーバ１４０３では、リクエスト受付手段１６１２がこのリクエストを受け付けてユーザリクエスト記憶部１６２３に記憶する（ステップＳ１８０２）。 First, the user operates the content request unit 1511 of the user terminal 1401 to request content from the web server 1403 (step S1801). In the web server 1403, the request receiving unit 1612 receives this request and stores it in the user request storage unit 1623 (step S1802).
　そして、コンテンツ配信手段１６１４がリクエストを受け付けたユーザの情報を読み出し手段１６１３に出力する（ステップＳ１８０３）。ここでいう「リクエストを受け付けたユーザの情報」とは、図１７のステップＳ１７０５で説明した通りである。そして読み出し手段１６１３は、コンテンツリクエスト記録部１６０２から、このユーザの過去のコンテンツ訪問履歴とコンテンツ訪問間隔を読み出して（ステップＳ１８０４）、これらを出力手段１６０６を通じて属性推定装置５００に入力する（ステップＳ１８０５）。 Then, the content distribution means 1614 outputs the information of the user who accepted the request to the reading means 1613 (step S1803). The “information of the user who accepted the request” here is as described in step S1705 of FIG. The reading unit 1613 reads the user's past content visit history and content visit interval from the content request recording unit 1602 (step S1804), and inputs them to the attribute estimation apparatus 500 through the output unit 1606 (step S1805). .
　属性推定装置５００では、尤度計算手段５１０が全属性モデルのパラメータ１１１を読み出し（ステップＳ１８０６）、これらの各属性モデルから、入力されたコンテンツ訪問履歴とコンテンツ訪問間隔に対する尤度を出力する（ステップＳ１８０７）。そして、各属性モデルから出力された尤度から、ユーザの属性を推定して、これを入力手段１６０１を通して広告選択手段１６１１に入力する（ステップＳ１８０８）。各属性モデルが出力した尤度の中から、最も高い尤度を出力した属性モデルの属性をそのユーザの推定結果とする。 In the attribute estimation device 500, the likelihood calculation means 510 reads the parameters 111 of all attribute models (step S1806), and outputs the likelihood for the input content visit history and content visit interval from each of these attribute models (step S1806). S1807). Then, the user's attribute is estimated from the likelihood output from each attribute model, and this is input to the advertisement selection means 1611 through the input means 1601 (step S1808). Among the likelihoods output by each attribute model, the attribute of the attribute model that outputs the highest likelihood is used as the estimation result of the user.
　ユーザの属性を入力された広告選択手段１６１１は、推定されたユーザの属性に対応した広告情報を広告記憶部１６２１から読み出し、この広告情報をコンテンツ配信手段１６１４に入力する（ステップＳ１８０９）。コンテンツ配信手段１６１４は、コンテンツ記憶部１６２２からユーザがリクエストしたコンテンツを読み出し、そこにこの広告情報を付け加えてユーザ端末１４０１に配信する（ステップＳ１８１０）。ユーザ端末１４０１では、コンテンツ表示手段１５１２がこの広告情報入りのコンテンツを受け取って表示する（ステップＳ１８１１）。 The advertisement selection unit 1611 to which the user attribute is input reads the advertisement information corresponding to the estimated user attribute from the advertisement storage unit 1621 and inputs this advertisement information to the content distribution unit 1614 (step S1809). The content distribution unit 1614 reads the content requested by the user from the content storage unit 1622, adds the advertisement information to the content, and distributes it to the user terminal 1401 (step S1810). In the user terminal 1401, the content display unit 1512 receives and displays the content including the advertisement information (step S1811).
　この実施の形態により、コンテンツ訪問履歴とコンテンツ訪問時間間隔から、ユーザの属性を適切に推定することで、ユーザの属性に応じた広告情報をユーザがリクエストしたコンテンツに付与できる。図１９は、図１８の動作によってユーザ端末１４０１に配信された、広告情報１９０２を含むコンテンツ１９０１の例を示す説明図である。 According to this embodiment, by appropriately estimating the user attribute from the content visit history and the content visit time interval, advertisement information corresponding to the user attribute can be given to the content requested by the user. FIG. 19 is an explanatory diagram showing an example of content 1901 including advertisement information 1902 distributed to the user terminal 1401 by the operation of FIG.
（第４の実施形態の変形例１）
　ここで、上述の第４の実施形態にはいくつかの変形が存在する。以下、それらの変形例について説明する。たとえば、第３の実施形態の変形例２と同様に、コンテンツ訪問履歴とコンテンツ訪問間隔に加えて、検索ワード履歴もユーザの属性推定に使用することもできる。 (Modification 1 of 4th Embodiment)
Here, there are some variations in the above-described fourth embodiment. Hereinafter, these modifications will be described. For example, as in the second modification of the third embodiment, in addition to the content visit history and the content visit interval, the search word history can also be used for user attribute estimation.
（第４の実施形態の変形例２）
　また、ユーザの属性に対応した広告情報の代わりに、ユーザの属性に対応した別のコンテンツへのリンク（ＵＲＬ）をコンテンツに付け加えて配信するようにすることもできる。 (Modification 2 of the fourth embodiment)
Further, instead of the advertisement information corresponding to the user attribute, a link (URL) to another content corresponding to the user attribute may be added to the content and distributed.
（第５の実施形態）
　本発明の第５の実施形態は、以上で説明した第４の実施形態に係るコンテンツ配信システムを、推定されたユーザの属性に応じて、ウェブサーバがそのユーザの属性に対応したページ構成でそのコンテンツを構成して配信するという実施形態である。 (Fifth embodiment)
According to the fifth embodiment of the present invention, the content distribution system according to the fourth embodiment described above has a page configuration in which the web server corresponds to the user attribute according to the estimated user attribute. This is an embodiment in which content is structured and distributed.
　より具体的には、前述の第４の実施形態に係るウェブサーバが、広告選択手段のかわりにユーザの属性に基づいてコンテンツのページ構成を選択するページ構成選択手段２０１２を有する。そして、コンテンツ配信手段１６１４がこのページ構成でコンテンツを構成してユーザ端末に配信する。 More specifically, the web server according to the above-described fourth embodiment includes a page configuration selection unit 2012 that selects a page configuration of content based on a user attribute instead of the advertisement selection unit. Then, the content distribution means 1614 configures the content with this page configuration and distributes it to the user terminal.
　この構成を備えることにより、本実施形態は、算出された各ユーザの属性に適する内容のページ構成でコンテンツをユーザ端末に配信することが可能となる。
　以下、これをより詳細に説明する。 By providing this configuration, the present embodiment can distribute the content to the user terminal with a page configuration that is suitable for the calculated attribute of each user.
Hereinafter, this will be described in more detail.
　本発明の第５の実施形態は、ユーザがコンテンツ（ウェブページ）をサーバにリクエストした際に、そのコンテンツのページ構成方法として、あらかじめ複数用意しておいたレイアウトのパターンからユーザの属性に応じたページ構成を特定し、そのページ構成によってコンテンツを構成して配信する。 In the fifth embodiment of the present invention, when a user requests a content (web page) from a server, a plurality of layout patterns prepared in advance correspond to the user's attributes as a method for configuring the content page. The page configuration is specified, and the content is configured and distributed by the page configuration.
　その際、ユーザの属性の推定に、第３の実施の形態に係る属性推定装置５００を利用する。この場合、推定されるユーザの属性は、主に性別、年齢、職業などである。これらの属性に応じて、より受け入れられやすいデザイン、レイアウトなどによるページ構成を使用した方が、より効果的であるからである。 At that time, the attribute estimation apparatus 500 according to the third embodiment is used to estimate the user's attribute. In this case, the estimated user attributes are mainly gender, age, occupation, and the like. This is because it is more effective to use a page structure based on a design, layout, etc. that are more acceptable depending on these attributes.
　本発明の第５の実施形態に係る各装置の接続構成は、図１４で説明した第４の実施形態に係る接続構成と同一である。ただし、ウェブサーバ１４０３は後述のウェブサーバ１４０３ｂに置き換えられる。また、ユーザ端末１４０１の構成は、図１５で説明した構成と同一である。 The connection configuration of each device according to the fifth embodiment of the present invention is the same as the connection configuration according to the fourth embodiment described in FIG. However, the web server 1403 is replaced with a web server 1403b described later. The configuration of the user terminal 1401 is the same as the configuration described in FIG.
　図２０は、本発明の第５の実施形態に係るウェブサーバ１４０３ｂの構成を示す説明図である。ウェブサーバ１４０３ｂは、図１６で示したウェブサーバ１４０３と比べて、広告選択手段１６１１の代わりにページ構成選択手段２０１２がＣＰＵ１６０２で実行される点と、ＨＤＤ１６０４で広告記憶部１６２１が省略されている点が相違点である。この点以外は、図１６で示したウェブサーバ１４０３と同一の動作を行う。 FIG. 20 is an explanatory diagram showing the configuration of the web server 1403b according to the fifth embodiment of the present invention. The web server 1403b is different from the web server 1403 shown in FIG. 16 in that the page configuration selection unit 2012 is executed by the CPU 1602 instead of the advertisement selection unit 1611 and the advertisement storage unit 1621 is omitted in the HDD 1604. Is the difference. Except this point, the same operation as the web server 1403 shown in FIG. 16 is performed.
　ウェブサーバ１４０３ｂは、ユーザがサーバにコンテンツをリクエストした際に、そのコンテンツのページ構成方法として、あらかじめ複数用意しておいたパターンからユーザの属性に応じたページ構成を特定し、そのページ構成によってコンテンツを配信する。コンテンツ配信手段１６１４は、ユーザからリクエストされたコンテンツをユーザの属性に適したページ構成によって配信する。 When the user requests content from the server, the web server 1403b specifies a page configuration corresponding to the user's attribute from a plurality of patterns prepared in advance as the content page configuration method, and the content is determined by the page configuration. To deliver. The content distribution unit 1614 distributes the content requested by the user with a page configuration suitable for the user attribute.
　ページ構成選択手段２０１２は、属性推定装置５００によるユーザの属性推定結果を入力手段１６０９から受け取り、コンテンツ記憶部１６０３から、ユーザの属性に応じたページ構成を取り出してコンテンツ配信手段１６１４に渡す。 The page configuration selection unit 2012 receives the user attribute estimation result from the attribute estimation apparatus 500 from the input unit 1609, extracts the page configuration corresponding to the user attribute from the content storage unit 1603, and passes it to the content distribution unit 1614.
　次に、本発明の第５の実施形態に係る動作について説明する。動作は、属性モデル学習処理と属性推定結果に対応したページ構成のコンテンツを配信する処理の２つに分かれる。属性モデル学習処理は、図１７で説明した第４の実施の形態に係る動作と同一である。 Next, an operation according to the fifth embodiment of the present invention will be described. The operation is divided into two processes, that is, an attribute model learning process and a process of delivering content having a page configuration corresponding to the attribute estimation result. The attribute model learning process is the same as the operation according to the fourth embodiment described with reference to FIG.
　図２１は、図１４、１５、２０に示したユーザ端末１４０１、ウェブサーバ１４０３ｂ、および属性推定装置５００が行う、属性推定結果に対応したページ構成のコンテンツを配信する処理の動作を示すシーケンス図である。ステップＳ１８０１～１８０８は、図１８で説明した第４の実施の形態に係る動作と同一であるので、同一の参照符号を付けている。この動作によって、尤度計算手段５１０がユーザの属性を推定し、これを入力手段１６０１を通してページ構成選択手段２０１２に出力する所までは完了する。 FIG. 21 is a sequence diagram illustrating an operation of a process of delivering content having a page configuration corresponding to the attribute estimation result performed by the user terminal 1401, the web server 1403b, and the attribute estimation apparatus 500 illustrated in FIGS. is there. Steps S1801 to 1808 are the same as those according to the fourth embodiment described with reference to FIG. With this operation, the likelihood calculation means 510 estimates the user's attributes and is completed until it is output to the page configuration selection means 2012 through the input means 1601.
　ユーザの属性を入力されたページ構成選択手段２０１２は、ユーザからリクエストを受けたコンテンツで、推定したユーザの属性に対応したページ構成をコンテンツ記憶部１６０３から読み出す（ステップＳ２１０９）。図２２および図２３は、同一のコンテンツで、ページ構成が異なる２つのコンテンツ２２０１および２３０１の例を示す説明図である。 The page configuration selection unit 2012 to which the user attribute is input reads out the page configuration corresponding to the estimated user attribute from the content storage unit 1603 with the content received from the user (step S2109). 22 and 23 are explanatory diagrams illustrating examples of two contents 2201 and 2301 having the same contents but different page configurations.
　ページ構成選択手段２０１２は読み出されたページ構成をコンテンツ配信手段１６１４に入力する。コンテンツ配信手段１６１４は、コンテンツ記憶部１６２２からユーザがリクエストしたコンテンツを読み出し、入力されたページ構成で構成した後、ユーザ端末１４０１に配信する（ステップＳ２１１０）。ユーザ端末１４０１では、コンテンツ表示手段１５１２がこのコンテンツを受け取って表示する（ステップＳ１８１１）。 The page configuration selection unit 2012 inputs the read page configuration to the content distribution unit 1614. The content distribution unit 1614 reads the content requested by the user from the content storage unit 1622, configures it with the input page configuration, and distributes it to the user terminal 1401 (step S2110). In the user terminal 1401, the content display unit 1512 receives and displays this content (step S1811).
　この実施形態により、ユーザがコンテンツをサーバにリクエストした際に、そのコンテンツのページ構成方法として、あらかじめ複数用意しておいたパターンからユーザの属性に応じたパターンを特定し、そのパターンで構成したコンテンツを配信できる。 According to this embodiment, when a user requests content from the server, as a page configuration method for the content, a pattern corresponding to the user's attribute is specified from a plurality of patterns prepared in advance, and the content is configured with the pattern. Can be delivered.
（第５の実施形態の変形例１）
　ここで、上述の第５の実施形態にはいくつかの変形が存在する。以下、それらの変形例について説明する。たとえば、第３および第４の実施形態の変形例２と同様に、コンテンツ訪問履歴とコンテンツ訪問間隔に加えて、検索ワード履歴もユーザの属性推定に使用することもできる。 (Modification 1 of 5th Embodiment)
Here, there are some variations in the fifth embodiment described above. Hereinafter, these modifications will be described. For example, as in the second modification of the third and fourth embodiments, in addition to the content visit history and the content visit interval, the search word history can also be used for user attribute estimation.
（第６の実施形態）
　本発明の第６の実施形態に係る尤度計算装置（なりすまし検知装置２４００）は、第１および第２の実施形態に係る尤度計算装置で、入力手段１０１が、ユーザの入力コマンドをシンボル、ユーザが一つの入力コマンドから次の入力コマンドを入力するまでの時間間隔を遷移時間間隔として各々入力を受け付ける。そしてこの装置が、計算された尤度に基づいてユーザがなりすましであるか否かを判断するなりすまし検知手段２４０２を有する。 (Sixth embodiment)
A likelihood calculation apparatus (spoofing detection apparatus 2400) according to the sixth embodiment of the present invention is a likelihood calculation apparatus according to the first and second embodiments, in which the input unit 101 uses a user input command as a symbol, The time interval from the time when the user inputs one input command to the next time input command is accepted as the transition time interval. And this apparatus has the impersonation detection means 2402 which judges whether a user is impersonating based on the calculated likelihood.
　この構成を備えることにより、本実施形態は、不正な方法でパスワードが盗まれてもこのパスワードを悪用してのなりすまし行為を検知することが可能となる。
　以下、これをより詳細に説明する。 By providing this configuration, the present embodiment can detect an impersonation act by misusing this password even if the password is stolen by an unauthorized method.
Hereinafter, this will be described in more detail.
　本発明の第６の実施形態は、以上で説明した第１および第２の実施形態に係る尤度計算装置を、あるユーザになりすましてシステムに侵入する「なりすまし」を検出するなりすまし検知装置として使用するという実施形態である。 In the sixth embodiment of the present invention, the likelihood calculation devices according to the first and second embodiments described above are used as an impersonation detection device that impersonates a user and detects “spoofing” that enters the system. This is an embodiment.
　ここでいう「なりすまし」とは、システムを操作する権限の無い者が、システムへ侵入して、本来は権限を持つ者以外見ることのできない機密情報を盗み出したり、悪事をはたらいてその者のせいにしたりすることである。多くの場合、なりすましは権限を持つ者のＩＤおよびパスワードを不正な方法で盗み出すことによって行われる。 “Spoofing” here means that someone who does not have the authority to operate the system intrudes into the system and steals confidential information that cannot be seen by anyone other than those who have authority. Or to do. In many cases, impersonation is performed by stealing an authorized person's ID and password in an unauthorized manner.
　そのようななりすましを発見する方法として、本来の権限を持つ者が普段から打ち込むコマンドやキータイピング等の入力履歴を保持して、普段のコマンド入力とは異なる入力が行われた場合に「なりすましである」と判定するという方法がある。そこで本実施形態では、本来の権限を持つ者が普段から打ち込むコマンドやキータイピング等の履歴を統計モデルで学習した後、入力されたコマンドに対して統計モデルの尤度を計算し、尤度を基になりすましかどうかを検知する。 As a method of discovering such impersonation, if an authorized person keeps an input history such as commands and key typing that are normally entered, and an input different from the usual command input is made, `` spoofing There is a method of determining "Yes". Therefore, in the present embodiment, after learning the history of commands and key typing that are normally entered by a person with the original authority using the statistical model, the likelihood of the statistical model is calculated for the input command, and the likelihood is calculated. Detects whether it is impersonating or not.
　図２４は、本発明の第６の実施形態に係るなりすまし検知装置２４００の構成を示す説明図である。なりすまし検知装置２４０は、図３に示した第２の実施形態に係る尤度計算装置３００と同様のコンピュータ装置であり、尤度計算装置３００と同様の入力手段１０１、ＣＰＵ１０２、ＲＡＭ１０３、ＨＤＤ１０４、出力手段１０５を備える。同一の呼称および参照番号で呼んでいる動作部の機能および動作については、第１～第３の実施形態として説明したものと同一である。 FIG. 24 is an explanatory diagram showing a configuration of an impersonation detection device 2400 according to the sixth embodiment of the present invention. The impersonation detection device 240 is a computer device similar to the likelihood calculation device 300 according to the second embodiment shown in FIG. 3, and has the same input means 101, CPU 102, RAM 103, HDD 104, output as the likelihood calculation device 300. Means 105 is provided. The functions and operations of the operation units called with the same names and reference numbers are the same as those described in the first to third embodiments.
　なりすまし検知装置２４００では、尤度計算手段１１０およびパラメータ推定手段３２０に加えてなりすまし検知手段２４０２が、コンピュータプログラムをＣＰＵ１０２が実行することによりソフトウェア上で構築される。 In the spoofing detection device 2400, the spoofing detection unit 2402 in addition to the likelihood calculation unit 110 and the parameter estimation unit 320 is constructed on software by the CPU 102 executing the computer program.
　入力手段１０１からこれらの動作手段に対して、前述のように、第１～３の実施形態と同様に、Ｎ個のシンボルＸとＮ個の遷移時間間隔Ｔとからなるデータ列（Ｘ、Ｔ）が入力される。Ｎ個のシンボルの列は、Ｘ＝｛Ｘ１、…、Ｘｎ｝、データ列Ｎ個の遷移時間間隔の列は、Ｔ＝｛Ｔ１、…、Ｔｎ｝である。 As described above, from the input unit 101 to these operation units, as in the first to third embodiments, a data string (X, T) composed of N symbols X and N transition time intervals T is used. ) Is entered. A sequence of N symbols is X = {X1,..., Xn}, and a sequence of N data transition time intervals is T = {T1,..., Tn}.
　図２５は、図２４に示したなりすまし検知装置２４００に入力されるデータ列（Ｘ、Ｔ）の例を示す表である。たとえば、Ｎ個のシンボルの列Ｘは、Ｌｉｎｕｘ（登録商標）のコマンドを１つのシンボルとしたコマンド履歴である。遷移時間間隔の列Ｔは、あるコマンドを入力した後に、次のコマンドを入力するまでの時間の列とする。 FIG. 25 is a table showing an example of a data string (X, T) input to the impersonation detection device 2400 shown in FIG. For example, a column X of N symbols is a command history having a Linux (registered trademark) command as one symbol. The transition time interval column T is a column of time until a next command is input after a certain command is input.
　尤度計算手段１０３は、入力手段１０１により入力された各データ列（Ｘ、Ｔ）に対して、モデルに対する尤度Ｐ（Ｘ，Τ）を計算し、計算結果を出力手段１０５に渡す。
パラメータ推定手段３２０は、入力手段１０１から入力されたデータ列（Ｘ、Ｔ）から、モデルのパラメータ１１１を推定し、このパラメータ１１１をＨＤＤ１０４に記憶する。ＨＤＤ１０４は、尤度計算手段に必要なパラメータ１１１を記憶する。 The likelihood calculation means 103 calculates the likelihood P (X, Τ) for the model for each data string (X, T) input by the input means 101 and passes the calculation result to the output means 105.
The parameter estimation unit 320 estimates the model parameter 111 from the data string (X, T) input from the input unit 101, and stores the parameter 111 in the HDD 104. The HDD 104 stores parameters 111 necessary for the likelihood calculation means.
　なりすまし検知手段２４０２は、尤度計算手段１０３で計算された尤度を基に、入力シーケンスがなりすましであるか否かを判断する。出力手段１０５は、なりすまし検知手段２４０２による判定結果を出力する。以上で説明した動作以外は、第１～３の実施形態として説明したものと同一である。 Impersonation detection means 2402 determines whether or not the input sequence is impersonation based on the likelihood calculated by the likelihood calculation means 103. The output unit 105 outputs the determination result by the spoofing detection unit 2402. The operations other than those described above are the same as those described as the first to third embodiments.
　図２６は、図２４で示した尤度計算手段１０３およびなりすまし検知手段２４０２が行う尤度の計算およびなりすましの検知の処理を表すフローチャートである。 FIG. 26 is a flowchart showing processing of likelihood calculation and spoofing detection performed by the likelihood calculation unit 103 and the spoofing detection unit 2402 shown in FIG.
　ステップＳ２０１～２０３は、図２に示した尤度計算の動作と同一である。この動作で尤度計算手段１０３が算出した尤度は、なりすまし検知手段２４０２に対して出力される。なりすまし検知手段２４０２は、この尤度から入力シーケンスがなりすましであるか否かを判断し（ステップＳ２６０１）、この判断結果を出力手段１０５を介して出力する（ステップＳ２６０２）。ステップＳ２６０１の動作は、たとえば尤度が閾値以上の場合になりすましであると判断することができる。 Steps S201 to S203 are the same as the likelihood calculation operation shown in FIG. The likelihood calculated by the likelihood calculating unit 103 in this operation is output to the spoofing detecting unit 2402. The impersonation detection unit 2402 determines whether or not the input sequence is impersonation from the likelihood (step S2601), and outputs the determination result via the output unit 105 (step S2602). The operation in step S2601 can be determined to be spoofing when, for example, the likelihood is greater than or equal to a threshold value.
　図２７は、図２６のステップＳ２６０２に示した処理で出力手段１０５を介して出力された判定データの例を示す説明図である。図２７に示した例では、たとえば閾値を「０．８」と予め設定してあり、１個目のデータ列は尤度「０．９」と閾値よりも高いのでなりすましであり、２個目のデータ列は尤度「０．３」と閾値よりも低いのでなりすましではないと判断している。 FIG. 27 is an explanatory diagram showing an example of determination data output via the output unit 105 in the process shown in step S2602 of FIG. In the example shown in FIG. 27, for example, the threshold value is preset as “0.8”, and the first data string is spoofed because the likelihood is “0.9”, which is higher than the threshold value. Since the likelihood of the data string is lower than the threshold of “0.3”, it is determined that the data string is not impersonated.
　パラメータ推定手段３２０がパラメータ１１１を推定する動作は、図４に示した第２の実施形態に係る動作と同一である。 The operation of the parameter estimation means 320 estimating the parameter 111 is the same as the operation according to the second embodiment shown in FIG.
　以上で説明した第６の実施形態により、例えばコマンドやキータイピング等から、その入力が本来の権限を持った者によって行われたか、なりすましによって行われたかを認識することができる。 According to the sixth embodiment described above, it is possible to recognize whether the input is performed by an authorized person or by impersonation from, for example, a command or key typing.
（第６の実施形態の変形例１）
　ここで、上述の第６の実施形態にはいくつかの変形が存在する。以下、それらの変形例について説明する。たとえばなりすまし検知装置２４００の構成を、図２４に示した構成から、パラメータ推定手段３２０を除いたものとしてもよい。図２８は、第６の実施形態の変形例１に係るなりすまし検知装置２４００ｂの構成を示す説明図である。この場合、パラメータ１１１は推定されるものではなく、予めＨＤＤ１０４に記憶されている。 (Modification 1 of 6th Embodiment)
Here, there are some variations in the sixth embodiment described above. Hereinafter, these modifications will be described. For example, the configuration of the spoofing detection device 2400 may be the configuration shown in FIG. 24 excluding the parameter estimation unit 320. FIG. 28 is an explanatory diagram illustrating a configuration of an impersonation detection device 2400b according to the first modification of the sixth embodiment. In this case, the parameter 111 is not estimated but is stored in the HDD 104 in advance.
（第６の実施形態の変形例２）
　また、なりすまし検知装置２４００に入力されるデータ列（Ｘ、Ｔ）で、コマンド時間間隔Ｔを離散値としてもよい。図２９は、コマンド時間間隔Ｔを離散値にしたデータ列（Ｘ、Ｔ）の例を示す説明図である。ここでは、コマンド時間間隔Ｔを、ＬＯＷ、ＨＩＧＨ、ＭＩＤの３段階に離散化している。 (Modification 2 of the sixth embodiment)
Further, the command time interval T may be a discrete value in the data string (X, T) input to the impersonation detection device 2400. FIG. 29 is an explanatory diagram showing an example of a data string (X, T) in which the command time interval T is a discrete value. Here, the command time interval T is discretized in three stages of LOW, HIGH, and MID.
（第６の実施形態の変形例３）
　さらに、なりすまし検知装置２４００に入力されるデータ列（Ｘ、Ｔ）で、入力データＸをコマンド単位ではなく１文字単位のキータイピング履歴としてもよい。図３０は、入力データＸをキータイピング履歴としたデータ列（Ｘ、Ｔ）の例を示す説明図である。 (Modification 3 of the sixth embodiment)
Further, in the data string (X, T) input to the impersonation detection device 2400, the input data X may be a key typing history for each character instead of for each command. FIG. 30 is an explanatory diagram illustrating an example of a data string (X, T) in which the input data X is a key typing history.
（第６の実施形態の変形例４）
　さらに、この変形例３で、変形例２と同様にコマンド時間間隔Ｔを離散値とすることもできる。図３１は、入力データＸをキータイピング履歴とし、さらにコマンド時間間隔ＴをＬＯＷ、ＨＩＧＨ、ＭＩＤの３段階の離散値としたデータ列（Ｘ、Ｔ）の例を示す説明図である。 (Modification 4 of the sixth embodiment)
Furthermore, in the third modification, the command time interval T can be a discrete value as in the second modification. FIG. 31 is an explanatory diagram showing an example of a data string (X, T) in which the input data X is a key typing history and the command time interval T is a discrete value of three levels of LOW, HIGH, and MID.
（第７の実施形態）
　本発明の第７の実施形態に係る尤度計算装置（動作認識装置３２００）は、第１および第２の実施形態に係る尤度計算装置で、入力手段１０１が、人物が写っている動画像中の各人物画像の特徴を示す特徴ベクトルをシンボル、特徴ベクトルが一つの状態から次の状態に移行するまでの時間間隔を遷移時間間隔として各々入力を受け付ける。そしてこの装置が、計算された尤度に基づいて人物画像に写っている人物が動いているか否かを判断する動作認識手段３２０２を有する。 (Seventh embodiment)
The likelihood calculation apparatus (motion recognition apparatus 3200) according to the seventh embodiment of the present invention is a likelihood calculation apparatus according to the first and second embodiments, and the input unit 101 is a moving image in which a person is shown. Each input is accepted with a feature vector indicating the feature of each human image as a symbol and a time interval until the feature vector transitions from one state to the next state as a transition time interval. And this apparatus has the action recognition means 3202 which judges whether the person reflected in the person image is moving based on the calculated likelihood.
　この構成を備えることにより、本実施形態は、画像に写っている人物が動いているか否かを判断することが可能となる。
　以下、これをより詳細に説明する。 With this configuration, the present embodiment can determine whether or not a person shown in the image is moving.
Hereinafter, this will be described in more detail.
　本発明の第７の実施形態は、以上で説明した第１および第２の実施形態に係る尤度計算装置を、人物が画像に納まった人物画像を時間順に並べた動画像から、人物の姿勢を認識する動作認識装置として使用するという実施形態である。ここでいう動作認識装置とは例えば、歩いている人の動画像を入力すると、歩いているという認識結果を出力し、歩いていない人の動画像を入力すると、歩いていないという認識結果を出力するものである。 In the seventh embodiment of the present invention, the likelihood calculation device according to the first and second embodiments described above is used to determine the posture of a person from a moving image in which person images arranged in the image are arranged in time order. It is embodiment using it as an operation | movement recognition apparatus which recognizes. For example, the motion recognition device here outputs a recognition result that a person is walking when a moving image of a person walking is input, and outputs a recognition result that a person is not walking when a moving image of a person who is not walking is input. To do.
　図３２は、本発明の第７の実施形態に係る動作認識装置３２００の構成を示す説明図である。動作認識装置３２００は、図３に示した第２の実施形態に係る尤度計算装置３００と同様のコンピュータ装置であり、尤度計算装置３００と同様の入力手段１０１、ＣＰＵ１０２、ＲＡＭ１０３、ＨＤＤ１０４、出力手段１０５を備える。同一の呼称および参照番号で呼んでいる動作部の機能および動作については、第１～第３の実施形態として説明したものと同一である。 FIG. 32 is an explanatory diagram showing the configuration of the motion recognition apparatus 3200 according to the seventh embodiment of the present invention. The motion recognition device 3200 is a computer device similar to the likelihood calculation device 300 according to the second embodiment shown in FIG. 3, and has the same input means 101, CPU 102, RAM 103, HDD 104, output as the likelihood calculation device 300. Means 105 is provided. The functions and operations of the operation units called with the same names and reference numbers are the same as those described in the first to third embodiments.
　動作認識装置３２００では、尤度計算手段１１０およびパラメータ推定手段３２０に加えて動作認識手段３２０２が、コンピュータプログラムをＣＰＵ１０２が実行することによりソフトウェア上で構築される。 In the motion recognition device 3200, in addition to the likelihood calculation means 110 and the parameter estimation means 320, the motion recognition means 3202 is constructed on software by the CPU 102 executing the computer program.
　入力手段１０１からこれらの動作手段に対して、前述のように、第１～３の実施形態と同様に、Ｎ個のシンボルＸとＮ個の遷移時間間隔Ｔとからなるデータ列（Ｘ、Ｔ）が入力される。Ｎ個のシンボルの列は、Ｘ＝｛Ｘ１、…、Ｘｎ｝、データ列Ｎ個の遷移時間間隔の列は、Ｔ＝｛Ｔ１、…、Ｔｎ｝である。 As described above, from the input unit 101 to these operation units, as in the first to third embodiments, a data string (X, T) composed of N symbols X and N transition time intervals T is used. ) Is entered. A sequence of N symbols is X = {X1,..., Xn}, and a sequence of N data transition time intervals is T = {T1,..., Tn}.
　ここでは、動画像を人物画像が並んだデータ列とみなし、データ列中のシンボルＸ列を、人物画像のＸ’，Ｙ，Ｚ軸の（Ｘ’，Ｙ，Ｚ）の列とする。人物領域の輪郭部の一点におけるｘ’，ｙ，，ｚを要素に持ち、全輪郭のｘ’，ｙ，，ｚをベクトルにまとめたものを、特徴ベクトル（Ｘ’，Ｙ，Ｚ）とする。図３３は、図３２に示した動作認識装置３２００に入力される特長ベクトルの例を示す説明図である。データ列Ｘは、特徴ベクトルの列となる。 Here, it is assumed that a moving image is a data sequence in which person images are arranged, and a symbol X column in the data sequence is an X ′, Y, Z axis (X ′, Y, Z) column of the person image. A feature vector (X ′, Y, Z) is obtained by combining x ′, y, z of all contours into a vector having x ′, y, z at one point of the contour portion of the human region as elements. . FIG. 33 is an explanatory diagram showing an example of a feature vector input to the motion recognition device 3200 shown in FIG. The data sequence X is a sequence of feature vectors.
　そして遷移時間間隔Ｔは、ある特徴ベクトルから次の特徴ベクトルに遷移する時間間隔を表す。撮像装置によって取得された特徴ベクトル（Ｘ’，Ｙ，Ｚ）を時刻順に並べた後に、２つの時刻の間で、特徴ベクトルの差がある閾値よりも大きい場合は特徴ベクトル（Ｘ’，Ｙ，Ｚ）を保存して、小さい場合は削除して保存しないこととする。よって、ある特徴ベクトルから次の特徴ベクトルに遷移する時間間隔は、人物の動きが大きく変化した間隔の特徴を表す。 The transition time interval T represents a time interval for transition from one feature vector to the next feature vector. After the feature vectors (X ′, Y, Z) acquired by the imaging device are arranged in order of time, if the difference between the feature vectors is greater than a certain threshold value between the two times, the feature vectors (X ′, Y, Z) is saved, and if it is small, it is not deleted and saved. Therefore, the time interval for transition from one feature vector to the next feature vector represents the feature of the interval at which the person's movement has changed significantly.
　図３４は、図３２に示した動作認識装置３２００に入力される入力データの例を示す説明図である。入力データは、図３４に示したように複数用意しても良い。図３４の例では、Ｍ個のデータ列（Ｘ、Ｔ）を入力する。 FIG. 34 is an explanatory diagram showing an example of input data input to the motion recognition device 3200 shown in FIG. A plurality of input data may be prepared as shown in FIG. In the example of FIG. 34, M data strings (X, T) are input.
　尤度計算手段１０３は、入力手段１０１により入力された各データ列（Ｘ、Ｔ）に対して、モデルのパラメータ１１１を用いて、モデルに対する尤度Ｐ（Ｘ，Τ）を計算し、計算結果を動作認識手段３２０２に渡す。パラメータ推定手段３２０は、入力手段１０１から入力されたＭ個のデータ列（Ｘ、Ｔ）から、モデルのパラメータ１１１を推定し、推定結果をＨＤＤ１０４に渡す。ＨＤＤ１０４は、このパラメータ１１１を記憶する。 The likelihood calculating means 103 calculates the likelihood P (X, Τ) for the model for each data string (X, T) input by the input means 101 using the model parameter 111, and the calculation result Is transferred to the motion recognition means 3202. The parameter estimation unit 320 estimates the model parameter 111 from the M data strings (X, T) input from the input unit 101 and passes the estimation result to the HDD 104. The HDD 104 stores the parameter 111.
　動作認識手段３２０２は、尤度計算手段１０３で計算された尤度を基に、入力シーケンスがモデルに当てはまる動作であるか否かを判断する。出力手段１０５は動作認識手段３２０２による判定結果を出力する。以上で説明した動作以外は、第１～３の実施形態として説明したものと同一である。 The motion recognition unit 3202 determines whether or not the input sequence is a motion that applies to the model based on the likelihood calculated by the likelihood calculation unit 103. The output unit 105 outputs the determination result by the motion recognition unit 3202. The operations other than those described above are the same as those described as the first to third embodiments.
　図３５は、図３２で示した尤度計算手段１０３および動作認識手段３２０２が行う尤度の計算および動作の認識の処理を表すフローチャートである。 FIG. 35 is a flowchart showing likelihood calculation and action recognition processing performed by the likelihood calculation means 103 and the action recognition means 3202 shown in FIG.
　ステップＳ２０１～２０３は、図２に示した尤度計算の動作と同一である。この動作で尤度計算手段１０３が算出した尤度は、動作認識手段３２０２に対して出力される。動作認識手段３２０２は、この尤度から入力シーケンスがモデルに当てはまる動作か否かを判断し（ステップＳ３５０１）、この判断結果を出力手段１０５を介して出力する（ステップＳ３５０２）。ステップＳ３５０１の動作は、たとえば尤度が閾値以上の場合にモデルに当てはまる動作であると判断することができる。 Steps S201 to S203 are the same as the likelihood calculation operation shown in FIG. The likelihood calculated by the likelihood calculating unit 103 in this operation is output to the operation recognizing unit 3202. The motion recognition unit 3202 determines whether or not the input sequence is a motion that applies to the model from the likelihood (step S3501), and outputs the determination result via the output unit 105 (step S3502). The operation in step S3501 can be determined to be an operation that applies to the model when the likelihood is equal to or greater than a threshold, for example.
　図３６は、図３５のステップＳ３５０２の動作で、出力手段１０５を介して出力された判定データの例を示す説明図である。図３６に示した例では、歩行動作のモデルに対して、たとえば閾値を「０．８」と予め設定してあり、１個目のデータ列は尤度「０．９」と閾値よりも高いので歩行をあらわすモデルに当てはまり、２個目のデータ列は、尤度「０．３」と閾値よりも低いので同モデルに当てはまらないと判断している。 FIG. 36 is an explanatory diagram illustrating an example of determination data output via the output unit 105 in the operation of step S3502 of FIG. In the example shown in FIG. 36, for example, a threshold value is preset as “0.8” for the walking motion model, and the first data string has a likelihood “0.9” which is higher than the threshold value. Therefore, it is applied to the model representing walking, and the second data string is lower than the threshold value with a likelihood “0.3”, so it is determined that it does not apply to the model.
　パラメータ推定手段３２０が、ある人物が１つの動作を行っているデータ列に対してパラメータ１１１を推定する動作は、図４に示した第２の実施形態に係る動作と同一である。 The operation in which the parameter estimation unit 320 estimates the parameter 111 for the data sequence in which a certain person performs one operation is the same as the operation according to the second embodiment shown in FIG.
　第７の実施の形態により、例えば、人物が歩行する動画像から生成したデータ列Ｘ、ＴをＭ個入力して、モデルのパラメータを推定する。その後、あるデータ列Ｘ、Ｔを入力すると、各データ列が、歩行をしているかそうでないかを認識することができる。 According to the seventh embodiment, for example, M data strings X and T generated from a moving image in which a person walks are input, and model parameters are estimated. Thereafter, when certain data strings X and T are input, it is possible to recognize whether each data string is walking or not.
（第７の実施形態の変形例１）
　ここで、上述の第７の実施形態にはいくつかの変形が存在する。以下、それらの変形例について説明する。たとえば、第７の実施の形態の動作認識装置３２００から、パラメータ推定手段３２０を除いた構成にしてもよい。図３７は、第７の実施形態の変形例１に係る動作認識装置３２００ｂの構成を示す説明図である。この場合、パラメータ１１１は推定されるものではなく、予めＨＤＤ１０４に記憶されている。 (Modification 1 of 7th Embodiment)
Here, there are some variations in the seventh embodiment described above. Hereinafter, these modifications will be described. For example, the parameter estimation unit 320 may be omitted from the motion recognition device 3200 according to the seventh embodiment. FIG. 37 is an explanatory diagram illustrating a configuration of an action recognition device 3200b according to Modification 1 of the seventh embodiment. In this case, the parameter 111 is not estimated but is stored in the HDD 104 in advance.
（第７の実施形態の変形例２）
　また、動作認識装置３２００または３２００ｂに入力されるデータ列（Ｘ、Ｔ）で、特徴ベクトル時間間隔Ｔを離散値とすることができる。 (Modification 2 of 7th Embodiment)
Further, the feature vector time interval T can be a discrete value in the data string (X, T) input to the motion recognition device 3200 or 3200b.
（第７の実施形態の変形例３）
　さらに、動作認識装置３２００で、カメラで撮影した人物画像から特徴ベクトルを抽出するのではなく、モーションキャプチャ等で取得した人物の各関節角度から特徴ベクトルを抽出してもよい。この場合、画像の座標Ｘ，Ｙ，Ｚを特徴とする代わりに、各軸周辺の回転運動角度を示すＲｏｌｌ，Ｐｉｔｃｈ，Ｙａｗを用いる。図３８は、このＲｏｌｌ，Ｐｉｔｃｈ，Ｙａｗを用いたデータ列（Ｘ、Ｔ）の例を示す説明図である。 (Modification 3 of 7th Embodiment)
Furthermore, the motion recognition device 3200 may extract feature vectors from the joint angles of a person acquired by motion capture or the like instead of extracting feature vectors from a person image taken by a camera. In this case, instead of characterizing the coordinates X, Y, and Z of the image, Roll, Pitch, and Yaw indicating the rotational motion angle around each axis are used. FIG. 38 is an explanatory diagram showing an example of a data string (X, T) using the Roll, Pitch, and Yaw.
（第７の実施形態の変形例４）
　さらに、この変形例４においても、変形例２と同様に特徴ベクトル時間間隔Ｔを離散値とすることができる。 (Modification 4 of 7th Embodiment)
Further, also in the fourth modification, the feature vector time interval T can be a discrete value as in the second modification.
　この出願は２００９年１月２６日に出願された日本出願特願２００９－０１４７０１を基礎とする優先権を主張し、その開示の全てをここに取り込む。 This application claims priority based on Japanese Patent Application No. 2009-014701 filed on Jan. 26, 2009, the entire disclosure of which is incorporated herein.
　これまで本発明について図面に示した特定の実施形態をもって説明してきたが、本発明は図面に示した実施形態に限定されるものではなく、本発明の効果を奏する限り、これまで知られたいかなる構成であっても採用することができる。 The present invention has been described with reference to the specific embodiments shown in the drawings. However, the present invention is not limited to the embodiments shown in the drawings, and any known hitherto provided that the effects of the present invention are achieved. Even if it is a structure, it is employable.
　本発明の具体的な適用例として、実施形態の中で、コンテンツ配信システムにおけるユーザ属性推定装置、なりすまし検知装置、動作認識装置といった例を示した。これら以外にも、シンボルと遷移時間間隔との組み合わせによって表現されるデータ系列に対して特定の統計モデルに対する尤度の計算を行う用途に対して、本発明は幅広く適用できる。 As specific application examples of the present invention, examples of a user attribute estimation device, an impersonation detection device, and an action recognition device in the content distribution system are shown in the embodiment. In addition to these, the present invention can be widely applied to the use of calculating the likelihood for a specific statistical model for a data series expressed by a combination of a symbol and a transition time interval.
　　１００、３００　尤度計算装置
　　１０１、１５０１、１６０１　入力手段
　　１０２、１５０２、１６０２　ＣＰＵ
　　１０３、１５０３、１６０３　ＲＡＭ
　　１０４、１６０４　ＨＤＤ
　　１０５、１５０５、１６０６　出力手段
　　１１０、５１０　尤度計算手段
　　１１１　パラメータ
　　３２０　パラメータ推定手段
　　５００　属性推定装置
　　５２０　属性モデル学習手段
　　１４００　コンテンツ配信システム
　　１４０１　ユーザ端末
　　１４０２　インターネット
　　１４０３、１４０３ｂ　ウェブサーバ
　　１５０４、１６０５　ネットワークカード
　　１５１１　コンテンツリクエスト手段
　　１５１２　コンテンツ表示手段
　　１６１１　広告選択手段
　　１６１２　リクエスト受付手段
　　１６１３　読み出し手段
　　１６１４　コンテンツ配信手段
　　１６２１　広告記憶部
　　１６２２　コンテンツ記憶部
　　１６２３　ユーザリクエスト記憶部
　　１９０１、２２０１、２３０１　コンテンツ
　　１９０２　広告情報
　　２０１２　ページ構成選択手段
　　２４００、２４００ｂ　なりすまし検知装置
　　２４０２　なりすまし検知手段
　　３２００、３２００ｂ　動作認識装置
　　３２０２　動作認識手段 100, 300 Likelihood calculation device 101, 1501, 1601 Input means 102, 1502, 1602 CPU
103, 1503, 1603 RAM
104, 1604 HDD
105, 1505, 1606 Output means 110, 510 Likelihood calculation means 111 Parameter 320 Parameter estimation means 500 Attribute estimation device 520 Attribute model learning means 1400 Content distribution system 1401 User terminal 1402 Internet 1403, 1403b Web server 1504, 1605 Network card 1511 Content Request unit 1512 Content display unit 1611 Advertisement selection unit 1612 Request reception unit 1613 Reading unit 1614 Content distribution unit 1621 Advertisement storage unit 1622 Content storage unit 1623 User request storage unit 1901, 21201, 3011, Content 1902 Advertisement information 2012 Page configuration selection unit 2400, 2400b Spoofing detection device 2402 Impersonation detection means 3200, 3200b Motion recognition device 3202 Motion recognition means

Claims (13)

  1. 　シンボルと遷移時間間隔との組み合わせによって表現されるデータ系列の入力を受け付ける入力手段と、
    　前記データ系列の特定の統計モデルを表すパラメータを記憶する記憶手段と、
    　前記パラメータを用いて前記データ系列の前記統計モデルに対する尤度を計算する尤度計算手段と、
    　この尤度計算手段によって計算された前記尤度を出力する出力手段を備えたことを特徴とする尤度計算装置。     An input means for receiving an input of a data series expressed by a combination of a symbol and a transition time interval;
    Storage means for storing parameters representing a particular statistical model of the data series;
    Likelihood calculating means for calculating the likelihood of the data series for the statistical model using the parameter;
    A likelihood calculating apparatus comprising output means for outputting the likelihood calculated by the likelihood calculating means.
  2. 　前記統計モデルが、隠れマルコフモデルまたはマルコフモデルのうちのいずれかによって表現され、前記遷移時間間隔が、離散値または連続値のうちのいずれかによって表現されることを特徴とする、請求項１に記載の尤度計算装置。 The method of claim 1, wherein the statistical model is represented by either a hidden Markov model or a Markov model, and the transition time interval is represented by either a discrete value or a continuous value. The likelihood calculation apparatus described.
  3. 　入力された前記データ系列を用いて前記統計モデルのパラメータを推定して前記記憶手段に記憶するパラメータ推定手段を有することを特徴とする、請求項１に記載の尤度計算装置。 The likelihood calculation apparatus according to claim 1, further comprising parameter estimation means for estimating the parameters of the statistical model using the inputted data series and storing them in the storage means.
  4. 　前記入力手段が、ユーザが訪問したウェブコンテンツの履歴を示すコンテンツ訪問履歴を前記シンボル、前記ユーザが一つのウェブコンテンツを訪問してから次のウェブコンテンツを訪問するまでのコンテンツ訪問間隔を前記遷移時間間隔として各々入力を受け付け、
    　前記尤度計算手段が、計算された前記尤度を前記ユーザの属性として出力することを特徴とする、請求項１ないし請求項３のうちいずれか１項に記載の尤度計算装置。     The input means is a symbol indicating a content visit history indicating a history of web content visited by a user, and a content visit interval from a visit of one web content to a visit of the next web content is the transition time. Accept each input as an interval,
    The likelihood calculation apparatus according to any one of claims 1 to 3, wherein the likelihood calculation means outputs the calculated likelihood as an attribute of the user.
  5. 　前記入力手段が入力を受け付ける前記シンボルが、前記ユーザが前記ウェブコンテンツを検索する際に使用した検索ワードの履歴を含むことを特徴とする、請求項４に記載の尤度計算装置。 The likelihood calculation apparatus according to claim 4, wherein the symbol that the input unit receives input includes a history of search words used when the user searches the web content.
  6. 　前記入力手段が、ユーザの入力コマンドを前記シンボル、前記ユーザが一つの入力コマンドから次の入力コマンドを入力するまでの時間間隔を前記遷移時間間隔として各々入力を受け付け、
    　前記尤度計算手段で計算された前記尤度に基づいて前記ユーザがなりすましであるか否かを判断するなりすまし検知手段を有することを特徴とする、請求項１ないし請求項３のうちいずれか１項に記載の尤度計算装置。     The input means accepts an input as a user input command as the symbol, and a time interval until the user inputs a next input command as the transition time interval,
    The spoofing detection means for judging whether or not the user is impersonating based on the likelihood calculated by the likelihood calculation means. The likelihood calculation apparatus according to the item.
  7. 　前記入力手段が、人物が写っている動画像中の各人物画像の特徴を示す特徴ベクトルを前記シンボル、前記特徴ベクトルが一つの状態から次の状態に移行するまでの時間間隔を前記遷移時間間隔として各々入力を受け付け、
    　前記尤度計算手段で計算された前記尤度に基づいて前記人物画像に写っている人物が動いているか否かを判断する動作認識手段を有することを特徴とする、請求項１ないし請求項３のうちいずれか１項に記載の尤度計算装置。     The input means has the symbol as a feature vector indicating a feature of each person image in a moving image in which a person is photographed, and a time interval until the feature vector shifts from one state to the next state as the transition time interval. Accept each input as
    4. The apparatus according to claim 1, further comprising motion recognition means for judging whether or not a person shown in the person image is moving based on the likelihood calculated by the likelihood calculation means. The likelihood calculation apparatus of any one of these.
  8. 　ユーザが操作するユーザ端末と、ウェブサーバと、尤度計算装置とがネットワークを介して相互に接続されたコンテンツ配信システムであって、
    　前記ウェブサーバが、前記ユーザ端末からのリクエストを受け付けるリクエスト受付手段と、前記リクエストに対応したコンテンツを前記ユーザ端末に配信するコンテンツ配信手段と、前記ユーザにリクエストされたコンテンツの履歴を示すコンテンツ訪問履歴と、前記ユーザが一つのコンテンツから次のコンテンツを訪問するまでのコンテンツ訪問間隔とを記憶するユーザリクエスト記憶部と、前記コンテンツ訪問履歴をシンボル、前記コンテンツ訪問間隔を遷移時間間隔とするデータ系列として前記尤度計算装置に入力する出力手段とを有し、
    　前記尤度計算装置が、前記シンボルと前記遷移時間間隔との組み合わせによって表現されるデータ系列の入力を受け付ける入力手段と、前記データ系列の特定の統計モデルを表すパラメータを記憶する記憶手段と、前記パラメータを用いて前記データ系列の前記統計モデルに対する尤度を計算する尤度計算手段とを有すると共に、この尤度計算手段によって計算された前記尤度を出力する出力手段を備え、
    　前記ウェブサーバが、前記尤度計算装置で計算された前記ユーザの属性の入力を受け付ける入力手段と、前記ユーザの属性に基づいて前記コンテンツに追加する広告を選択する広告選択手段とを有し、前記コンテンツ配信手段が前記コンテンツに前記広告を追加して前記ユーザ端末に配信する機能を持つことを特徴とするコンテンツ配信システム。     A content distribution system in which a user terminal operated by a user, a web server, and a likelihood calculation device are connected to each other via a network,
    Request reception means for receiving a request from the user terminal, content distribution means for distributing content corresponding to the request to the user terminal, and a content visit history indicating a history of content requested by the user A user request storage unit that stores a content visit interval until the user visits the next content from one content, and a data sequence in which the content visit history is a symbol and the content visit interval is a transition time interval Output means for inputting to the likelihood calculation device,
    The likelihood calculating device includes an input unit that receives an input of a data series expressed by a combination of the symbol and the transition time interval, a storage unit that stores a parameter representing a specific statistical model of the data series, and And a likelihood calculating means for calculating the likelihood of the data series for the statistical model using a parameter, and an output means for outputting the likelihood calculated by the likelihood calculating means,
    The web server includes input means for receiving input of the user attribute calculated by the likelihood calculating device, and advertisement selection means for selecting an advertisement to be added to the content based on the user attribute; The content distribution system, wherein the content distribution unit has a function of adding the advertisement to the content and distributing it to the user terminal.
  9. 　ユーザが操作するユーザ端末と、ウェブサーバと、尤度計算装置とがネットワークを介して相互に接続されたコンテンツ配信システムであって、
    　前記ウェブサーバが、前記ユーザ端末からのリクエストを受け付けるリクエスト受付手段と、前記リクエストに対応したコンテンツを前記ユーザ端末に配信するコンテンツ配信手段と、前記ユーザにリクエストされたコンテンツの履歴を示すコンテンツ訪問履歴と、前記ユーザが一つのコンテンツから次のコンテンツを訪問するまでのコンテンツ訪問間隔とを記憶するユーザリクエスト記憶部と、前記コンテンツ訪問履歴をシンボル、前記コンテンツ訪問間隔を遷移時間間隔とするデータ系列として前記尤度計算装置に入力する出力手段とを有し、
    　前記尤度計算装置が、前記シンボルと前記遷移時間間隔との組み合わせによって表現されるデータ系列の入力を受け付ける入力手段と、前記データ系列の特定の統計モデルを表すパラメータを記憶する記憶手段と、前記パラメータを用いて前記データ系列の前記統計モデルに対する尤度を計算する尤度計算手段とを有すると共に、この尤度計算手段によって計算された前記尤度を出力する出力手段を備え、
    　前記ウェブサーバが、前記尤度計算装置で計算された前記ユーザの属性の入力を受け付ける入力手段と、前記ユーザの属性に基づいて前記コンテンツのページ構成を選択するページ構成選択手段とを有し、前記コンテンツ配信手段が前記ページ構成によって前記コンテンツを前記ユーザ端末に配信する機能を持つことを特徴とするコンテンツ配信システム。     A content distribution system in which a user terminal operated by a user, a web server, and a likelihood calculation device are connected to each other via a network,
    Request reception means for receiving a request from the user terminal, content distribution means for distributing content corresponding to the request to the user terminal, and a content visit history indicating a history of content requested by the user A user request storage unit that stores a content visit interval until the user visits the next content from one content, and a data sequence in which the content visit history is a symbol and the content visit interval is a transition time interval Output means for input to the likelihood calculation device,
    The likelihood calculating device includes an input unit that receives an input of a data series expressed by a combination of the symbol and the transition time interval, a storage unit that stores a parameter representing a specific statistical model of the data series, and And a likelihood calculating means for calculating the likelihood of the data series for the statistical model using a parameter, and an output means for outputting the likelihood calculated by the likelihood calculating means,
    The web server has input means for receiving input of the user's attribute calculated by the likelihood calculating device, and page configuration selection means for selecting a page configuration of the content based on the user's attribute; The content distribution system, wherein the content distribution means has a function of distributing the content to the user terminal according to the page configuration.
  10. 　シンボルと遷移時間間隔との組み合わせによって表現されるデータ系列に対して、前記データ系列の特定の統計モデルに対する尤度を計算する尤度計算方法であって、
    　入力手段が前記データ系列の入力を受け付け、
    　記憶手段から前記統計モデルのパラメータを読み出し、
    　前記パラメータを用いて尤度計算手段が前記データ系列の前記統計モデルに対する尤度を計算し、
    　計算された前記尤度を出力手段に出力することを特徴とする尤度計算方法。     A likelihood calculation method for calculating a likelihood for a specific statistical model of the data series for a data series expressed by a combination of a symbol and a transition time interval,
    The input means accepts the input of the data series,
    Read the parameters of the statistical model from the storage means,
    Using the parameter, likelihood calculating means calculates the likelihood of the data series for the statistical model,
    A likelihood calculation method characterized by outputting the calculated likelihood to an output means.
  11. 　入力された前記データ系列を用いてパラメータ推定手段が前記統計モデルのパラメータを推定し、
    　推定された前記パラメータを前記パラメータ推定手段が前記記憶手段に記憶することを特徴とする、請求項１０に記載の尤度計算方法。     Parameter estimation means estimates the parameters of the statistical model using the input data series,
    The likelihood calculation method according to claim 10, wherein the parameter estimation unit stores the estimated parameter in the storage unit.
  12. 　シンボルと遷移時間間隔との組み合わせによって表現されるデータ系列に対して、前記データ系列の特定の統計モデルに対する尤度を計算する尤度計算装置であって、
    　前記データ系列の入力を受け付ける手順と、
    　前記統計モデルのパラメータを読み出す手順と、
    　前記パラメータを用いて前記データ系列の前記統計モデルに対する尤度を計算する手順と、
    　計算された前記尤度を出力手段に出力する手順と
    をコンピュータに実行させることを特徴とする尤度計算プログラム。     A likelihood calculating device that calculates the likelihood of a specific statistical model of the data series for a data series expressed by a combination of a symbol and a transition time interval,
    A procedure for receiving input of the data series;
    A procedure for reading the parameters of the statistical model;
    Calculating a likelihood of the data series for the statistical model using the parameters;
    A likelihood calculation program that causes a computer to execute a procedure for outputting the calculated likelihood to an output means.
  13. 　入力された前記データ系列を用いて前記統計モデルのパラメータを推定する手順と、
    　推定された前記パラメータをする手順と記憶する手順と
    をコンピュータに実行させることを特徴とする、請求項１２に記載の尤度計算プログラム。     Estimating the parameters of the statistical model using the input data series;
    The likelihood calculation program according to claim 12, wherein the computer executes a procedure for performing the estimated parameter and a procedure for storing the parameter.
PCT/JP2010/050493 2009-01-26 2010-01-18 Likelihood estimation device, content delivery system, likelihood estimation method, and likelihood estimation program WO2010084839A1 (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
JP2010547477A JPWO2010084839A1 (en) 2009-01-26 2010-01-18 Likelihood estimation apparatus, content distribution system, likelihood estimation method, and likelihood estimation program

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
JP2009-014701 2009-01-26
JP2009014701 2009-01-26

Publications (1)

Publication Number Publication Date
WO2010084839A1 true WO2010084839A1 (en) 2010-07-29

Family

ID=42355893

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/JP2010/050493 WO2010084839A1 (en) 2009-01-26 2010-01-18 Likelihood estimation device, content delivery system, likelihood estimation method, and likelihood estimation program

Country Status (2)

Country Link
JP (1) JPWO2010084839A1 (en)
WO (1) WO2010084839A1 (en)

Cited By (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
EP2361991A1 (en) 2002-12-18 2011-08-31 Third Wave Technologies, Inc. Detection of small nucleic acids
JP2015018491A (en) * 2013-07-12 2015-01-29 ヤフー株式会社 Information processing device and method
JP2016045543A (en) * 2014-08-20 2016-04-04 ヤフー株式会社 Information processing apparatus, information processing method, and program
JP2017134750A (en) * 2016-01-29 2017-08-03 ヤフー株式会社 Authentication device, authentication method and authentication program
JP2018077765A (en) * 2016-11-11 2018-05-17 ヤフー株式会社 Information processing device, information processing method, and information processing program
JP2018156393A (en) * 2017-03-17 2018-10-04 ヤフー株式会社 Estimating apparatus, estimating method, and estimating program
JP2021068328A (en) * 2019-10-25 2021-04-30 ヤフー株式会社 Information providing device, information providing method, and information providing program
JP7034459B1 (en) 2021-06-14 2022-03-14 株式会社Insight science Information processing equipment, information processing systems, and programs

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JPH11259643A (en) * 1998-01-30 1999-09-24 Mitsubishi Electric Inf Technol Center America Inc Action detection system
JP2006107473A (en) * 2004-10-05 2006-04-20 Microsoft Corp System, method, and interface for providing personalized search and information access
WO2007075236A1 (en) * 2005-12-19 2007-07-05 Microsoft Corporation Delivery of personalized keyword-based information using client-side re-ranking

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JPH11259643A (en) * 1998-01-30 1999-09-24 Mitsubishi Electric Inf Technol Center America Inc Action detection system
JP2006107473A (en) * 2004-10-05 2006-04-20 Microsoft Corp System, method, and interface for providing personalized search and information access
WO2007075236A1 (en) * 2005-12-19 2007-07-05 Microsoft Corporation Delivery of personalized keyword-based information using client-side re-ranking

Non-Patent Citations (3)

* Cited by examiner, † Cited by third party
Title
KENJI YAMANISHI ET AL.: "Joho Security to AI Data Mining no Joho Security eno Oyo", JOURNAL OF JAPANESE SOCIETY FOR ARTIFICIAL INTELLIGENCE, vol. 21, no. 5, 1 September 2006 (2006-09-01), pages 570 - 576 *
OSAMU YOSHIDA ET AL.: "Web Access Log ni Taisuru Keiretsu Data Mining -Page Taizai Jikan Keiretsu no Kaiseki", IEICE TECHNICAL REPORT, vol. 106, no. 340, 3 November 2006 (2006-11-03), pages 13 - 18 *
YUKI KOSAKA ET AL.: "Web Access Sen'i ni Motozuita Yuryo Kokyakudo Scoring", DAI 7 KAI FORUM ON INFORMATION TECHNOLOGY KOEN RONBUNSHU SEPARATE, vol. 2, 20 August 2008 (2008-08-20), pages 423 - 424 *

Cited By (10)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
EP2361991A1 (en) 2002-12-18 2011-08-31 Third Wave Technologies, Inc. Detection of small nucleic acids
JP2015018491A (en) * 2013-07-12 2015-01-29 ヤフー株式会社 Information processing device and method
JP2016045543A (en) * 2014-08-20 2016-04-04 ヤフー株式会社 Information processing apparatus, information processing method, and program
JP2017134750A (en) * 2016-01-29 2017-08-03 ヤフー株式会社 Authentication device, authentication method and authentication program
JP2018077765A (en) * 2016-11-11 2018-05-17 ヤフー株式会社 Information processing device, information processing method, and information processing program
JP2018156393A (en) * 2017-03-17 2018-10-04 ヤフー株式会社 Estimating apparatus, estimating method, and estimating program
JP2021068328A (en) * 2019-10-25 2021-04-30 ヤフー株式会社 Information providing device, information providing method, and information providing program
JP7072549B2 (en) 2019-10-25 2022-05-20 ヤフー株式会社 Information providing equipment, information providing method and information providing program
JP7034459B1 (en) 2021-06-14 2022-03-14 株式会社Insight science Information processing equipment, information processing systems, and programs
JP2022190645A (en) * 2021-06-14 2022-12-26 株式会社Insight science Information processing apparatus, information processing system, and program

Also Published As

Publication number Publication date
JPWO2010084839A1 (en) 2012-07-19

Similar Documents

Publication Publication Date Title
WO2010084839A1 (en) Likelihood estimation device, content delivery system, likelihood estimation method, and likelihood estimation program
Yuan et al. Video summarization by learning deep side semantic embedding
WO2022041979A1 (en) Information recommendation model training method and related device
US10783361B2 (en) Predictive analysis of target behaviors utilizing RNN-based user embeddings
US11443170B2 (en) Semi-supervised training of neural networks
Zhou et al. Property inference attacks against GANs
US9569536B2 (en) Identifying similar applications
CN109729395B (en) Video quality evaluation method and device, storage medium and computer equipment
US20160188726A1 (en) Scalable user intent mining using a multimodal restricted boltzmann machine
CN106462608A (en) Knowledge source personalization to improve language models
Olmezogullari et al. Pattern2Vec: Representation of clickstream data sequences for learning user navigational behavior
CN105144227A (en) Social media impact assessment
US20110054999A1 (en) System and method for predicting user navigation within sponsored search advertisements
US9344507B2 (en) Method of processing web access information and server implementing same
CN108846091A (en) Information recommended method, device and equipment
CN110197389A (en) A kind of user identification method and device
KR102523160B1 (en) Method and system for detecting duplicated document using document similarity measuring model based on deep learning
CN113886697A (en) Clustering algorithm based activity recommendation method, device, equipment and storage medium
CN114943549A (en) Advertisement delivery method and device
JP6989434B2 (en) Programs, devices and methods for estimating potential target users who will reach the predicted target behavior using the access history of the URL.
CN111127057B (en) Multi-dimensional user portrait recovery method
CN110717094A (en) Information recommendation method and device, computer equipment and storage medium
CN111967924A (en) Commodity recommendation method, commodity recommendation device, computer device, and medium
Shi et al. Predicting the next scenic spot a user will browse on a tourism website based on markov prediction model
JP5588938B2 (en) Item recommendation apparatus, method and program

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 10733446

Country of ref document: EP

Kind code of ref document: A1

DPE1 Request for preliminary examination filed after expiration of 19th month from priority date (pct application filed from 20040101)
ENP Entry into the national phase

Ref document number: 2010547477

Country of ref document: JP

Kind code of ref document: A

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 10733446

Country of ref document: EP

Kind code of ref document: A1