WO2017041541A1 - 推送推荐信息的方法、服务器及存储介质 - Google Patents

推送推荐信息的方法、服务器及存储介质 Download PDF

Info

Publication number
WO2017041541A1
WO2017041541A1 PCT/CN2016/084284 CN2016084284W WO2017041541A1 WO 2017041541 A1 WO2017041541 A1 WO 2017041541A1 CN 2016084284 W CN2016084284 W CN 2016084284W WO 2017041541 A1 WO2017041541 A1 WO 2017041541A1
Authority
WO
WIPO (PCT)
Prior art keywords
user
candidate
path
attribute value
meta
Prior art date
Application number
PCT/CN2016/084284
Other languages
English (en)
French (fr)
Inventor
石川
贺鹏
易玲玲
张志强
Original Assignee
北京邮电大学
腾讯科技(深圳)有限公司
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 北京邮电大学, 腾讯科技(深圳)有限公司 filed Critical 北京邮电大学
Publication of WO2017041541A1 publication Critical patent/WO2017041541A1/zh
Priority to US15/715,840 priority Critical patent/US10609433B2/en

Links

Images

Classifications

    • HELECTRICITY
    • H04ELECTRIC COMMUNICATION TECHNIQUE
    • H04NPICTORIAL COMMUNICATION, e.g. TELEVISION
    • H04N21/00Selective content distribution, e.g. interactive television or video on demand [VOD]
    • H04N21/20Servers specifically adapted for the distribution of content, e.g. VOD servers; Operations thereof
    • H04N21/25Management operations performed by the server for facilitating the content distribution or administrating data related to end-users or client devices, e.g. end-user or client device authentication, learning user preferences for recommending movies
    • H04N21/251Learning process for intelligent management, e.g. learning user preferences for recommending movies
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00Information retrieval; Database structures therefor; File system structures therefor
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/90Details of database functions independent of the retrieved data types
    • G06F16/95Retrieval from the web
    • G06F16/953Querying, e.g. by the use of web search engines
    • G06F16/9535Search customisation based on user profiles and personalisation
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/90Details of database functions independent of the retrieved data types
    • G06F16/95Retrieval from the web
    • G06F16/958Organisation or management of web site content, e.g. publishing, maintaining pages or automatic linking
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06QINFORMATION 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
    • G06Q30/00Commerce
    • G06Q30/02Marketing; Price estimation or determination; Fundraising
    • G06Q30/0241Advertisements
    • G06Q30/0251Targeted advertisements
    • G06Q30/0255Targeted advertisements based on user history
    • HELECTRICITY
    • H04ELECTRIC COMMUNICATION TECHNIQUE
    • H04NPICTORIAL COMMUNICATION, e.g. TELEVISION
    • H04N21/00Selective content distribution, e.g. interactive television or video on demand [VOD]
    • H04N21/20Servers specifically adapted for the distribution of content, e.g. VOD servers; Operations thereof
    • H04N21/25Management operations performed by the server for facilitating the content distribution or administrating data related to end-users or client devices, e.g. end-user or client device authentication, learning user preferences for recommending movies
    • H04N21/251Learning process for intelligent management, e.g. learning user preferences for recommending movies
    • H04N21/252Processing of multiple end-users' preferences to derive collaborative data
    • HELECTRICITY
    • H04ELECTRIC COMMUNICATION TECHNIQUE
    • H04NPICTORIAL COMMUNICATION, e.g. TELEVISION
    • H04N21/00Selective content distribution, e.g. interactive television or video on demand [VOD]
    • H04N21/20Servers specifically adapted for the distribution of content, e.g. VOD servers; Operations thereof
    • H04N21/25Management operations performed by the server for facilitating the content distribution or administrating data related to end-users or client devices, e.g. end-user or client device authentication, learning user preferences for recommending movies
    • H04N21/258Client or end-user data management, e.g. managing client capabilities, user preferences or demographics, processing of multiple end-users preferences to derive collaborative data
    • H04N21/25866Management of end-user data
    • H04N21/25891Management of end-user data being end-user preferences
    • HELECTRICITY
    • H04ELECTRIC COMMUNICATION TECHNIQUE
    • H04NPICTORIAL COMMUNICATION, e.g. TELEVISION
    • H04N21/00Selective content distribution, e.g. interactive television or video on demand [VOD]
    • H04N21/40Client devices specifically adapted for the reception of or interaction with content, e.g. set-top-box [STB]; Operations thereof
    • H04N21/43Processing of content or additional data, e.g. demultiplexing additional data from a digital video stream; Elementary client operations, e.g. monitoring of home network or synchronising decoder's clock; Client middleware
    • H04N21/442Monitoring of processes or resources, e.g. detecting the failure of a recording device, monitoring the downstream bandwidth, the number of times a movie has been viewed, the storage space available from the internal hard disk
    • H04N21/44213Monitoring of end-user related data
    • H04N21/44222Analytics of user selections, e.g. selection of programs or purchase activity
    • H04N21/44224Monitoring of user activity on external systems, e.g. Internet browsing
    • H04N21/44226Monitoring of user activity on external systems, e.g. Internet browsing on social networks
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06QINFORMATION 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
    • G06Q30/00Commerce
    • G06Q30/06Buying, selling or leasing transactions
    • G06Q30/0601Electronic shopping [e-shopping]
    • G06Q30/0631Item recommendations

Definitions

  • the present invention relates to the field of information processing technologies, and in particular, to a method, a server, and a storage medium for pushing recommended information.
  • a current method for pushing recommendation information is mainly implemented based on a user's social relationship. For example, if user A watches movie M and user A and user B are friends, then movie M is recommended to user B.
  • the current method of pushing recommendation information only considers social relationships between users, but users with social relationships do not necessarily have the same recommendation requirements. For example, user A and user B are friend relationships, but user A and user B may have completely different viewing preferences, so that it is inaccurate to recommend the movie M that user A has watched to user B. It can be seen that the current recommendation method of pushing recommendation information based on social relationship is not accurate and needs no improvement.
  • Embodiments of the present invention provide a method, server, and storage medium for pushing recommendation information.
  • a method of pushing recommendation information including:
  • the meta path package a connection with the attribute value between the candidate user and the candidate recommendation object;
  • the recommendation information of the candidate recommendation object is sent to the terminal corresponding to the target user.
  • a server comprising a memory and a processor, wherein the memory stores instructions, wherein when the instructions are executed by the processor, the processor performs the following steps:
  • the meta path includes a connection with the attribute value between the candidate user and the candidate recommended object
  • the recommendation information of the candidate recommendation object is sent to the terminal corresponding to the target user.
  • One or more non-volatile readable storage media storing computer-executable instructions, when executed by one or more processors, cause the one or more processors to perform the following steps:
  • the meta path includes a connection with the attribute value between the candidate user and the candidate recommended object
  • FIG. 1 is a schematic diagram of a network mode of a heterogeneous information network for recommending a movie in an embodiment
  • FIG. 2 is a schematic diagram of a heterogeneous information network with attribute values for recommending movies in one embodiment
  • FIG. 3 is an application environment diagram of a recommendation system in an embodiment
  • FIG. 4 is a schematic structural diagram of a server for implementing a method for pushing recommendation information in an embodiment
  • FIG. 5 is a schematic flow chart of a method for pushing recommendation information in an embodiment
  • FIG. 6 is a schematic flowchart of a step of obtaining a user similarity between a target user and a candidate user with respect to a meta path in an embodiment
  • FIG. 7 is a comparison diagram of calculating a user similarity under a meta path without attribute values and a meta path with attribute values in one embodiment
  • FIG. 8 is a flow chart showing the steps of estimating the attribute value of the connection between the candidate recommended object and the target user according to the attribute value of the connection between the candidate user and the candidate recommended object, the attribute value constraint condition of the meta path, and the user similarity in one embodiment.
  • FIG. 9 is a structural block diagram of an apparatus for pushing recommendation information in an embodiment
  • Figure 10 is a block diagram showing the structure of an apparatus for pushing recommended information in another embodiment.
  • the heterogeneous information network is first The related concepts are explained.
  • the underlying data structure of the heterogeneous information network is a directed graph.
  • the heterogeneous information network contains different types of objects and connections, and the connections between the objects represent different relationships. Heterogeneity and rich information make heterogeneous information networks a better form of data representation in many scenarios.
  • the meta-path is a unique feature in heterogeneous information networks. It connects two objects in the network through a sequence of relationships between two object types, and is widely used to explore rich semantic information.
  • the attributes of the user, user attributes, candidate recommendation objects, and candidate recommendation objects can all be objects in the heterogeneous information network.
  • a connection represents a relationship between connected objects.
  • Heterogeneous information networks can be used to recommend movies.
  • the heterogeneous information network includes different types of objects, such as users and movies. It also includes various relationships between these objects, such as viewing information, social networks, and objects. Attribute information. Heterogeneous information networks can effectively integrate a variety of information that may be useful for recommendations. In addition, exploring the different semantics of objects and relationships in the network can reveal subtle relationships between objects.
  • the meta-path "user-movie-user" in Figure 1 represents a user who has seen the same movie, and this meta-path can be used for similar user retrieval based on the viewing record. If a movie is recommended based on this meta-path, it will recommend movies that have been viewed by other users who have the same viewing history as the target user. Similarly, a user with similar interests can be found based on the meta-path "user-interest group-user". Therefore, we can find similar users according to the different meta paths of the connected users, and then directly recommend the recommended objects that these similar users like. Furthermore, different recommendation models can be implemented by reasonably setting different meta paths.
  • heterogeneous information networks and meta-paths do not take into account attribute values on the connection, but heterogeneous information networks for recommending movies may contain connections with attribute values.
  • the user will give a score of between 1 and 5 for the movie he has seen (as shown in the connection between the user and the movie in Fig. 1), and the higher the rating value indicates that the user likes the movie. Similar user searches may result in inaccurate results if the score size is not considered. For example, based on the path "user-movie-user", Tom's similarity to Mary or Bob is the same because they all watched the same movie. However, they may give different ratings for completely different interests. In fact, Tom and Bob score very high on the same movie, so based on the score, they are Like.
  • object type or relationship type And attribute value type
  • the network can be referred to as a heterogeneous information network without attribute values.
  • object type (or relationship type And attribute value type The network can be referred to as a heterogeneous information network with attribute values.
  • a conventional heterogeneous information network has no attribute values, that is, relationships in the network have no attribute values or are not considered.
  • some relationships in the network have attribute values that may be continuous or discrete. Continuous attribute values can be converted to discrete attribute value processing.
  • the user can score a score between 1 and 5 for the movie that has been viewed; in the heterogeneous information network of the scientific literature, the relationship between the author and the paper can be taken. Different attribute values indicate that the author is the first author of the paper.
  • a heterogeneous information network with attribute values for recommending movies is provided, and Figure 2 shows its network mode.
  • This heterogeneous information network includes objects of six different object types (such as users, movies, interest groups, actors, directors, and movie types) and the relationships between them.
  • the connections between objects represent different relationships. For example, a connection between users represents a social relationship, and a connection between a user and a movie represents a rating relationship.
  • the rating relationship between the user and the movie also has an attribute value ranging from 1 to 5 integers.
  • Two objects in a heterogeneous information network can be connected by different meta paths, and these paths have different meanings.
  • users can connect through a path such as "user-user (UU)", “user-group-user” (UGU for short), and "user-movie-user” (UMU for short).
  • UU user-user
  • UGU user-group-user
  • UU user-movie-user
  • the meta-path refers to the meta-path with the attribute value constraint on the connection, which can be expressed as (can also be abbreviated as The subscript l indicates the number of the meta path.
  • the attribute value function ⁇ (Re) is a set of attribute values on the relationship Re, otherwise ⁇ (Re) is an empty set.
  • the relationship Re l between A l and A l+1 is based on the attribute value ⁇ l (Re l ).
  • Attribute value constraints on attribute value functions Is a collection of associated constraints between attribute value functions. If all attribute value functions on the meta path are empty sets (corresponding attribute value constraints) It is also an empty set), then such a meta path is called a meta path without attribute values, otherwise it is called a meta path with an attribute value or an extended meta path.
  • the score value of the scoring relationship between the user U and the movie M may take a value of 1-5.
  • Meta path ie U(1)M indicates that the user rated the movie as 1 point, which implies that the user does not like the movie.
  • Meta path ie, U(1,2)M(1,2)U indicates a candidate user who does not like the same movie as the target user, and the meta-path UMU without the attribute value can only reflect the user with the same movie record.
  • U(1,2)M(1,2)U indicates a candidate user who does not like the same movie as the target user, and the meta-path UMU without the attribute value can only reflect the user with the same movie record.
  • Bo is similar, but they are not similar to Mary.
  • a recommendation system includes a server 302 and a terminal 304 connected by a network.
  • the server 302 can be an independent physical server or a server cluster composed of multiple physical servers.
  • the terminal 304 may be a desktop computer or a mobile terminal including at least one of a mobile phone, a smart watch, a tablet, and a PDA (Personal Digital Assistant).
  • server 302 includes a connection through a system bus.
  • Processors non-volatile storage media, internal storage, and network interfaces.
  • the processor has a computing function and a function to control the operation of the server 302, the processor being configured to perform a method of pushing the recommendation information.
  • the non-volatile storage medium includes at least one of a magnetic storage medium, an optical storage medium, and a flash storage medium.
  • the non-volatile storage medium stores an operating system and means for pushing recommended information, and the means for pushing the recommended information is used to implement a method of pushing recommended information.
  • the network interface is used to connect to the network to communicate with the terminal 304.
  • a method for pushing recommendation information is provided. This embodiment is exemplified by the method applied to the server 302 in FIG. 3 and FIG. 4 described above.
  • Step 502 Acquire a meta path connecting the candidate user and the target user in the heterogeneous information network; the meta path includes a connection with the attribute value between the candidate user and the candidate recommended object.
  • the heterogeneous information network includes various types of objects including at least a candidate user, a target user, and a candidate recommended object, and the connection between the objects represents a relationship between the connected objects.
  • One of the users here is a data object to which the natural person is mapped
  • the target user represents the recipient of the recommendation information
  • the candidate user is the user whose relationship and attribute value are known.
  • the candidate recommendation object refers to an object that can be recommended to the target user, including at least one of a movie, a music, a book, a friend, a group, and an item.
  • the heterogeneous information network includes a meta-path that connects candidate users and target users, and includes candidate recommendation objects.
  • the object type of the meta path is symmetric.
  • the meta path may be “user-group-user”, “user-movie-user”, “user-movie- At least one of a movie type - movie - user”, "user-user-user”, and "user-movie-director-movie-user”.
  • Step 504 Acquire user similarity of the target user and the candidate user with respect to the meta path.
  • Similarity is a measure of similarity that indicates how similar two objects are.
  • the user similarity of the target user and the candidate user with respect to the meta path is based on the similarity between the target user and the candidate user calculated by the meta path. Since a long meta-path is meaningless and produces a bad similarity, the length of the meta-path can be limited to no more than four. The length of the meta path is equal to the number of connections in the meta path.
  • step 504 includes: comparing the target user to the candidate user relative to the meta The similarity of each atomic path of the path to obtain the user similarity of the target user and the candidate user with respect to the meta path.
  • the attribute value function ⁇ l (Re l ) takes a fixed value, so such a path is called an atomic meta path.
  • a meta path is a collection of all atomic path paths that satisfy the attribute value constraints of the meta path.
  • existing similarity measures can be used directly.
  • Existing similarity measures such as PathSim (Y.Sun, J. Han, X. Yan, P. Yu, and T. Wu. Pathsim: Meta path-based top-k similarity search in heterogeneous information networks. In VLDB, Pages 992-1003, 2011), PCRW (N.Lao and W.
  • Step 506 Estimate the attribute value of the connection between the candidate recommendation object and the target user according to the attribute value of the connection between the candidate user and the candidate recommendation object, the attribute value constraint condition of the meta path, and the user similarity.
  • the meta path further includes an attribute value constraint for constraining the attribute value between the candidate user and the candidate recommended object on the meta path, and the attribute value of the connection between the candidate recommended object and the target user.
  • the attribute value constraint may be that the two attribute values are equal or differ within a preset range.
  • the two attribute values are equal, which is a strict attribute value constraint, which makes the recommendation result more accurate.
  • a broader meta-path semantics can be mined.
  • the attribute value i and the attribute value j must satisfy the attribute value constraint:
  • the user similarity may reflect the degree of similarity between the candidate user and the target user, and thus may be used to determine the degree to which the attribute value corresponding to the corresponding candidate user is considered when estimating the attribute value of the connection between the candidate recommended object and the target user.
  • the attribute value constraint can be used to define a specific value of the estimated attribute value such that the estimated attribute value conforms to the semantics expressed by the meta path.
  • Step 508 When the estimated attribute value satisfies the recommendation condition, the recommendation information of the candidate recommendation object is sent to the terminal corresponding to the target user.
  • the recommendation condition is whether to recommend the judgment condition of the corresponding candidate recommendation object to the target user, and the recommendation condition may be, for example, that the estimated attribute value is greater than a preset threshold, or the estimated attribute value is equal to the preset threshold or the estimated attribute value is less than the pre-predetermined value.
  • the threshold which is determined according to the meaning of the attribute value and the recommended requirements.
  • the attribute value indicates the user's rating value for the movie
  • the rating value is positively related to the user's attitude toward the movie, that is, the more the user likes the movie rating value
  • the estimated attribute value is greater than or equal to the preset.
  • the recommendation information of the candidate recommended object is transmitted to the terminal corresponding to the target user. If the rating value is negatively related to the user's attitude towards the movie, that is, the user prefers the lower the movie rating value, then when the estimated attribute value is less than the preset threshold, the candidate recommendation object is sent to the terminal corresponding to the target user.
  • Recommended information can be flexibly configured according to the recommended accuracy requirements.
  • the recommendation information may include description information of the candidate recommendation object, and may also include an access address of the candidate recommendation object.
  • the description information may include information such as a movie name, a movie overview, a director, an actor, and a poster
  • the access address may be an access address of the ticket purchase website or an access address of the online video website.
  • the method for pushing the recommendation information is implemented by a novel heterogeneous information network, which includes a meta-path connecting the candidate user and the target user, and can represent the social relationship between the target user and the candidate user.
  • the connection between the candidate user and the candidate recommended object in the meta path has an attribute value to quantify the relationship between the candidate user and the candidate recommended object.
  • the candidate recommendation object and the target may be estimated according to the user similarity combined with the attribute value of the connection between the candidate user and the candidate recommended object and the attribute value constraint condition.
  • the attribute value of the connection between users, the estimated attribute value can be reversed A quantified relationship between the target user and the candidate recommendation object is reflected. In this way, when the heterogeneous information network is used for recommendation, not only the social relationship of the target user but also the quantitative relationship between the target user and the candidate recommended object is considered, so that the push result is more accurate.
  • step 504 specifically includes the following steps:
  • Step 602 Split the meta path into a plurality of atomic path according to the attribute value constraint condition of the meta path.
  • U(1)M(1)U and U(1)M(2)U are both atomic paths.
  • Step 604 Obtain a similarity between the target user and the candidate user with respect to each atomic path.
  • any similarity measure methods such as PathSim, PCRW, and HeteSim may be used to calculate the similarity between the target user and the candidate user with respect to each atomic path.
  • PathSim the number of path instances connecting the target user and the candidate user is calculated along the atomic path first, and then the quantity is regularized to obtain the corresponding similarity.
  • Step 606 Calculate user similarity of the target user and the candidate user with respect to the meta path according to the acquired similarity with respect to each atomic path.
  • the similarity with respect to each atomic path is based on the similarity of each atomic path. Since the meta-path can be split into a set of corresponding atomic paths, the user similarity based on the meta-path can be regarded as a comprehensive similarity based on the similarity of all corresponding atomic paths. User similarity can be obtained by summing or weighting.
  • step 606 includes calculating the acquired relative to each atomic path.
  • the sum of similarities; the sum of similarities is directly or positively correlated as the user similarity of the target user and the candidate user with respect to the meta path.
  • the similarities of all the atomic paths of the meta path are summed.
  • the sum of these similarities can then be directly used as the user similarity of the target user and the candidate user with respect to the meta path, or the positive correlation operation between the similarities can be used as the target user and the candidate user is similar to the user of the meta path. degree.
  • the positive correlation operation refers to an operation in which the dependent variable is consistent with the trend of the independent variable, such as adding or subtracting or multiplying or dividing by a positive value.
  • the positive correlation operation includes regularization processing.
  • the user similarity calculated by the two similarity measures of PathSim and HeteSim needs to be regularized to the user similarity to limit the range of similarity calculated.
  • a process of calculating user similarity is exemplified by a heterogeneous information network for recommending a movie.
  • the users u 1 , u 2 , and u 3 both watch the movie m 1 and the movie m 2 and give corresponding score values, such as the score value matrix in Fig. 7.
  • the similarity matrix calculated by PathSim represents u 1 , u 2 .
  • the similarities between u and 2 are equal.
  • the PathSim can be directly used to calculate u 1 , u 2 and u on each atomic path. 3 similarity between two and two.
  • different paths atoms membered u 1, u 2 and u 3 of similarity between any two of the rules are summed and can be obtained membered path U (i) M (j) U
  • i j under similar Degree matrix. Only the user can see the user u 1 and u 2 are similar, since user u 1 and u 2 users have the same movie viewing preferences.
  • step 506 specifically includes the following steps:
  • Step 802 Obtain a discrete value range of attribute values of the connection between the target user and the candidate recommendation object.
  • R represents a real number set.
  • Step 804 For each value in the discrete value range, respectively obtain a connection between the candidate user and the candidate recommendation object with the attribute value whose value satisfies the attribute value constraint condition, and the candidate user corresponding to the obtained connection and The user similarity between the target users calculates the attribute value strength corresponding to the value.
  • is a user similarity matrix, indicating the similarity between two users in the user set U, where Indicates the target user u and the candidate user v relative to the meta path User similarity.
  • ⁇ N is defined, where Indicated at a given path
  • the attribute value of the connection of the lower target user u and the candidate recommended object x is the strength of r.
  • E v,x,r indicates whether the candidate user v has a property value of r with the candidate recommendation object x, and if so, E v,x,r is 1, otherwise 0.
  • the attribute value constraint is used for the case where the two attribute values are equal.
  • E v, x, r can be modified correspondingly only in R v, x and r satisfy the attribute value constraint condition.
  • the time value is 1.
  • Step 806 Calculate the weighted average value by using each of the values in the discrete value range as the weight of the corresponding attribute value.
  • each value r of 1 to N by the corresponding attribute value strength Weighting is performed to calculate a weighted average.
  • the attribute value strength can also be After the regularization process, each of the values in the discrete value range is calculated as a weighted average with the corresponding regularized attribute value strength as a weight. Attribute value strength The regularization process is specifically: using the attribute value strength Divides by the sum of the strengths of all attribute values under the corresponding meta path.
  • Step 808 Obtain an estimated attribute value of the connection between the candidate recommendation object and the target user according to the calculated weighted average value.
  • the weighted average calculated in step 806 can be directly used as the estimated attribute value.
  • the specific estimated attribute value can be calculated by the following formula (2):
  • N represents the upper limit of the discrete value range.
  • equation (2) has the added advantage that it can eliminate the bias of user similarity calculated in different meta-paths. Considering that the user similarity calculated based on different meta-paths has different value ranges, this makes it difficult to compare the similarity calculation and the attribute value strength between different meta-paths, and the regularized attribute value strength in formula (2) can be eliminated. The difference in the range of values.
  • step 808 includes multiplying the weighted average values calculated under the respective meta-paths by the path weights of the corresponding meta-paths to calculate a weighted average value, and obtaining an estimate of the connection between the candidate recommended object and the target user. Property value.
  • a unified path weight is set for each user path for each user, indicating the user's preference for the metadata path. Specifically as formula (3):
  • w (l) represents the meta path Path weight.
  • a comprehensive estimated attribute value based on all meta paths Can use the estimated attribute value on each meta path
  • the weighted average is expressed.
  • the sum of the path weights of the respective meta-paths obtained after the target optimization is 1, so That is, the weighted average calculated under each meta path is multiplied by the path weight of the corresponding meta path to calculate a weighted average.
  • the symbol ⁇ represents the Adama product of the matrix, that is, the product of the corresponding elements;
  • p represents the p-norm of the matrix.
  • St represents the constraint.
  • the path weight vector can be solved by optimizing the objective function of the above formula (4).
  • step 808 includes: multiplying the weighted average values calculated under the respective meta-paths by path weights corresponding to the target users and the corresponding meta-paths to calculate a weighted average value, and obtaining between the candidate recommended objects and the target users.
  • the estimated attribute value of the connection includes: multiplying the weighted average values calculated under the respective meta-paths by path weights corresponding to the target users and the corresponding meta-paths to calculate a weighted average value, and obtaining between the candidate recommended objects and the target users. The estimated attribute value of the connection.
  • denotes the Hadamard product of the matrix, i.e., the product of the corresponding element
  • p represents p matrix norm.
  • the path weight matrix W can be solved by optimizing the objective function of the above formula (6).
  • the method for pushing the recommendation information further includes: obtaining a real attribute value of the connection between the candidate recommendation object and the target user; initializing the path weight corresponding to the target user and the meta path; according to the user similarity, The initialized path weight is adjusted in a direction approaching the average of the path weights corresponding to the candidate user and the meta path, such that the difference between the real attribute value and the estimated attribute value satisfies the minimum condition.
  • formula (6) takes into account the user-defined path weights, it is difficult for users who have only a small amount of attribute value information to perform effective weight learning.
  • the weight of learning needs to be shared The number of samples trained is much smaller than
  • the user's path weight should be consistent with the path weight of its similar users. For users with only a small number of attribute values, their path weights can be learned from the path weights of other users who are similar to them, because the user similarity based on the meta path is more effective for these users.
  • the path weight corresponding to the target user and the meta path is initialized, and may be initialized to a value of 0 or greater than 0, and then the initialized path weight is adjusted toward the direction of the average value of the path weight corresponding to the candidate user and the meta path.
  • the speed of approach is positively related to the similarity of the user. The greater the user similarity, the faster the approach, and the lower the user similarity, the slower the approach.
  • the adjustment is stopped when the difference in the value satisfies the minimum condition.
  • the minimum condition can be the above formula (4) or (6) or the following formula (9).
  • the path weight regularization term can be expressed in the form of a matrix of the following formula (8):
  • W (l) is the path weight matrix
  • 2 represents the 2 norm of the matrix.
  • represents the product of the matrix of the Adama, that is, the product of the corresponding elements
  • p represents the p-norm of the matrix
  • Y represents an indicator matrix
  • the following path weight vector; diag(W (l) ) represents the diagonal matrix transformed by the vector W (l) ; W represents the path weight matrix of all users; Estimated based meta-path
  • the attribute value matrix, R represents the true attribute value matrix.
  • the above formula (9) is a non-negative quadratic programming problem, that is, a simple form of non-negative matrix factorization, which can be solved by using the gradient projection method for solving the problem with non-negative bound constrained optimization problems, and solving the non-negative boundary
  • the gradient projection method of the constrained optimization problem can be referred to "CJ Lin. Projected gradient methods for non-negative matrix factorization. In Neural Computation, pages 2756-2279, 2007".
  • Formula (9) The gradient is:
  • is the step size and can be set as needed.
  • Steps (1) to (7) may be referred to as a Semantic path based personalized Recommendation method, and a semantic path based personalized recommendation method.
  • Step (1) acquiring a heterogeneous information network G with attribute values and a meta path set of the connected user Control parameter ⁇ 0, the control parameters ⁇ 1, and the convergence step size ⁇ when the threshold value ⁇ parameter update.
  • Step (2) relative to the meta path set In each of the meta-paths, calculate the user similarity matrix S (l) , the attribute value strength matrix Q (l), and the estimated attribute value matrix
  • step (3) the path weight matrix W>0 is initialized.
  • step (7) the path weight matrix W of all users is output.
  • W means assigning W to W old .
  • the candidate recommendation object is a network resource; the attribute value is a rating value.
  • the network resources include resources that can be obtained from the network, such as movies, audio, and novels.
  • the score value can be used to reflect the user's quantitative attitude toward network resources.
  • the first data set contains 13367 user object users, 12677 movies, and 1068278 1-5 ratings.
  • the first data set also contains the social relationships of the user object users and the attribute information of the user object users and movies.
  • the second data set contains the rating values of the user object user to the local merchant, as well as the attribute information of the user object user and the merchant.
  • This data set contains 16,239 user object users, 14,284 local merchants, and 1983 97 1-5 rating values.
  • Table 1 is the detailed statistics of the two data sets.
  • the two datasets have some different properties.
  • the scores of the first dataset are more dense but the social relationship is very sparse, while the scores of the second dataset are sparse but more socially intensive.
  • RMSE average root mean square error
  • MAE mean absolute error
  • R test represents the entire test set.
  • a data set is divided into a training set and a test set.
  • the training set is used to train a heterogeneous information network with attribute values
  • the test set is used to test the effect of the trained heterogeneous information network.
  • a smaller MAE or RMSE indicates a better effect.
  • SemRec Reg personalized path weight learning method with path weighting regularizations
  • SemRec Sgl a method based on a single-element path
  • SemRec All a method of unifying path weights
  • SemRec Ind a method of learning the weight of personalized paths for each user
  • Second data set UGU UU U(i)M(j)U
  • i j UCoU U(i)MDM(j)U
  • i j U(i)B(j)U
  • i j U(i)MAM(j)U
  • i j U(i)BCaB(j)U
  • i j U(i)MTM(j)U
  • i j U(i)BCiB(j)U
  • i j
  • the training data ratio of 20% indicates that 20% of the scores in the user-candidate recommendation target score value matrix are used as training sets for model training, and the remaining 80% of the scores are predicted.
  • the first data set has a more dense rating value relationship.
  • the score of the second data set is more sparse, so more data is used as the training set on the second data set (60%, 70%, 80%, 90%).
  • 10 training sets and test sets were randomly and randomly divided according to a given ratio, and the average value was used as the result shown in Table 3.
  • SemRec (such as SemRec All and SemRec Reg ) of multiple paths has better effect than SemRec (SemRec Sgl ) of single path, except SemRec Ind , which means that SemRec's path weight learning method can be effectively integrated. Similarity information generated on different paths. Due to the sparsity of the score value, SemRec Ind is more effective than SemRec All in most cases. In addition, SemRec Reg achieves the best results in all situations. This is because SemRec Reg not only implements personalized path weight learning for all users, but also uses path weight regularization to avoid the problems caused by sparse scores.
  • SemRec Sgl and SemRec All are very fast and can be applied directly to online learning.
  • SemRec Ind and SemRec Reg runtimes are also acceptable. In real-world applications, you can choose a suitable SemRec method to balance efficiency and performance as needed.
  • an apparatus 900 for pushing recommendation information having functional modules for implementing the method of pushing recommendation information of the various embodiments described above, is provided.
  • the apparatus 900 for pushing recommendation information includes: a meta path acquisition module 901, a user similarity acquisition module 902, an attribute value estimation module 903, and a push module 904.
  • the meta path obtaining module 901 is configured to acquire a meta path connecting the candidate user and the target user in the heterogeneous information network; the meta path includes a connection with the attribute value between the candidate user and the candidate recommended object.
  • the user similarity obtaining module 902 is configured to acquire the user similarity of the target user and the candidate user with respect to the meta path.
  • the attribute value estimation module 903 is configured to estimate an attribute value of the connection between the candidate recommendation object and the target user according to the attribute value of the connection between the candidate user and the candidate recommendation object, the attribute value constraint condition of the meta path, and the user similarity.
  • the pushing module 904 is configured to: when the estimated attribute value meets the recommended condition, send the recommendation information of the candidate recommended object to the terminal corresponding to the target user.
  • the user similarity obtaining module 902 includes: split mode Block 902a, similarity calculation module 902b, and user similarity synthesis module 902c.
  • the splitting module 902a is configured to split the meta path into a plurality of atomic path according to the attribute value constraint condition of the meta path.
  • the similarity calculation module 902b is configured to obtain the similarity between the target user and the candidate user with respect to each atomic path.
  • the user similarity synthesis module 902c is configured to calculate a user similarity of the target user and the candidate user with respect to the meta path according to the obtained similarity with respect to each atomic element path.
  • the user similarity synthesis module 902c is further configured to calculate a sum of similarities obtained with respect to each atomic path; compare the sum of similarities directly or positively as a target user and a candidate user. The user similarity of the Yuyuan path.
  • the attribute value estimation module 903 includes a discrete value range acquisition module 903a, an attribute value strength calculation module 903b, a weighted average module 903c, and an estimation result generation module 903d.
  • the discrete value range obtaining module 903a is configured to obtain a discrete value range of the attribute values of the connection between the target user and the candidate recommended object.
  • the attribute value strength calculation module 903b is configured to obtain, for each value in the discrete value range, a connection between the candidate user and the candidate recommendation object having the attribute value that satisfies the attribute value constraint condition, according to the obtained connection.
  • the user similarity between the corresponding candidate user and the target user calculates the attribute value strength corresponding to the value.
  • the weighted average module 903c is configured to calculate a weighted average value by using each of the values in the discrete value range as the weight of the corresponding attribute value.
  • the estimation result generation module 903d is configured to obtain an estimated attribute value of the connection between the candidate recommendation object and the target user according to the calculated weighted average value.
  • the estimation result generating module 903d is further configured to multiply the weighted average values calculated under the respective meta paths by the path weights of the corresponding meta paths to calculate a weighted average value, and obtain a candidate recommendation object and a target user.
  • the estimated attribute value of the connection is further configured to multiply the weighted average values calculated under the respective meta paths by the path weights of the corresponding meta paths to calculate a weighted average value, and obtain a candidate recommendation object and a target user.
  • the estimation result generating module 903d is further configured to calculate the respective meta paths.
  • the weighted average values are respectively multiplied by the path weights corresponding to the target users and the corresponding meta paths to calculate a weighted average value, and the estimated attribute values of the connections between the candidate recommended objects and the target users are obtained.
  • the apparatus 900 for pushing recommendation information further includes: a path weight learning module 905, configured to acquire a real attribute value of a connection between the candidate recommended object and the target user; and a path weight corresponding to the target user and the meta path Initialization; according to the user similarity, the initialized path weight is adjusted toward the direction of the average value of the path weights corresponding to the candidate user and the meta-path, so that the difference between the real attribute value and the estimated attribute value satisfies the minimum condition.
  • a path weight learning module 905 configured to acquire a real attribute value of a connection between the candidate recommended object and the target user
  • a path weight corresponding to the target user and the meta path Initialization according to the user similarity, the initialized path weight is adjusted toward the direction of the average value of the path weights corresponding to the candidate user and the meta-path, so that the difference between the real attribute value and the estimated attribute value satisfies the minimum condition.
  • the candidate recommendation object is a network resource; the attribute value is a rating value.
  • the storage medium may be a non-volatile storage medium such as a magnetic disk, an optical disk, a read-only memory (ROM), or a random access memory (RAM).

Landscapes

  • Engineering & Computer Science (AREA)
  • Databases & Information Systems (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Business, Economics & Management (AREA)
  • Multimedia (AREA)
  • Signal Processing (AREA)
  • General Engineering & Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Accounting & Taxation (AREA)
  • Strategic Management (AREA)
  • Development Economics (AREA)
  • Finance (AREA)
  • Computing Systems (AREA)
  • General Business, Economics & Management (AREA)
  • Game Theory and Decision Science (AREA)
  • Entrepreneurship & Innovation (AREA)
  • Marketing (AREA)
  • Economics (AREA)
  • Health & Medical Sciences (AREA)
  • General Health & Medical Sciences (AREA)
  • Social Psychology (AREA)
  • Computer Graphics (AREA)
  • Computer Networks & Wireless Communication (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

一种推送推荐信息的方法,所述方法包括:获取异质信息网络中连接候选用户和目标用户的元路径;所述元路径包括所述候选用户和候选推荐对象间的具有属性值的连接;获取所述目标用户和所述候选用户相对于所述元路径的用户相似度;根据所述候选用户和候选推荐对象间的连接的属性值、所述元路径的属性值约束条件以及所述用户相似度,估计所述候选推荐对象和所述目标用户间的连接的属性值;当估计的属性值满足推荐条件时,向所述目标用户对应的终端发送所述候选推荐对象的推荐信息。

Description

推送推荐信息的方法、服务器及存储介质
本申请要求于2015年9月8日提交中国专利局,申请号为201510567428.9,发明名称为“推送推荐信息的方法和装置”的中国专利申请的优先权,其全部内容通过引用结合在本申请中。
技术领域
本发明涉及信息处理技术领域,特别是涉及一种推送推荐信息的方法、服务器及存储介质。
背景技术
通过向用户候选推荐对象,比如向用户推荐电影、音乐、书籍、好友、群组或者商品,使得用户无需主动搜索就可以获得相应被候选推荐对象的信息,为用户被动获取信息提供了一种途径。目前的一种推送推荐信息的方法主要是基于用户的社交关系来实现的,比如若用户A观看了电影M,且用户A与用户B是好友关系,那么就会将电影M推荐给用户B。
然而,目前的推送推荐信息的方法仅考虑了用户之间的社交关系,但具有社交关系的用户之间并不一定具有相同的推荐需求。比如用户A和用户B是好友关系,但用户A和用户B可能有完全不同的观影偏好,这样将用户A观看过的电影M推荐给用户B就是不准确的。可见,目前基于社交关系的推送推荐信息的方法的推荐结果并不准确,亟须改进。
发明内容
本发明的实施例提供一种推送推荐信息的方法、服务器及存储介质。
一种推送推荐信息的方法,包括:
获取异质信息网络中连接候选用户和目标用户的元路径;所述元路径包 括所述候选用户和候选推荐对象间的具有属性值的连接;
获取所述目标用户和所述候选用户相对于所述元路径的用户相似度;
根据所述候选用户和候选推荐对象间的连接的属性值、所述元路径的属性值约束条件以及所述用户相似度,估计所述候选推荐对象和所述目标用户间的连接的属性值;及
当估计的属性值满足推荐条件时,向所述目标用户对应的终端发送所述候选推荐对象的推荐信息。
一种服务器,包括存储器和处理器,所述存储器中储存有指令,其特征在于,所述指令被所述处理器执行时,使得所述处理器执行以下步骤:
获取异质信息网络中连接候选用户和目标用户的元路径;所述元路径包括所述候选用户和候选推荐对象间的具有属性值的连接;
获取所述目标用户和所述候选用户相对于所述元路径的用户相似度;
根据所述候选用户和候选推荐对象间的连接的属性值、所述元路径的属性值约束条件以及所述用户相似度,估计所述候选推荐对象和所述目标用户间的连接的属性值;及
当估计的属性值满足推荐条件时,向所述目标用户对应的终端发送所述候选推荐对象的推荐信息。
一个或多个存储有计算机可执行指令的非易失性可读存储介质,所述计算机可执行指令被一个或多个处理器执行时,使得所述一个或多个处理器执行以下步骤:
获取异质信息网络中连接候选用户和目标用户的元路径;所述元路径包括所述候选用户和候选推荐对象间的具有属性值的连接;
获取所述目标用户和所述候选用户相对于所述元路径的用户相似度;
根据所述候选用户和候选推荐对象间的连接的属性值、所述元路径的属性值约束条件以及所述用户相似度,估计所述候选推荐对象和所述目标用户间的连接的属性值;及
当估计的属性值满足推荐条件时,向所述目标用户对应的终端发送所述 候选推荐对象的推荐信息。
本发明的一个或多个实施例的细节在下面的附图和描述中提出。本发明的其它特征、目的和优点将从说明书、附图以及权利要求书变得明显。
附图说明
图1为一个实施例中用于推荐电影的异质信息网络的网络模式示意图;
图2为一个实施例中用于推荐电影的带属性值的异质信息网络的示意图;
图3为一个实施例中推荐***的应用环境图;
图4为一个实施例中用于实现推送推荐信息的方法的服务器的结构示意图;
图5为一个实施例中推送推荐信息的方法的流程示意图;
图6为一个实施例中获取目标用户和候选用户相对于元路径的用户相似度的步骤的流程示意图;
图7为一个实施例中不带属性值的元路径和带属性值的元路径下计算用户相似度的对比图;
图8为一个实施例中根据候选用户和候选推荐对象间的连接的属性值、元路径的属性值约束条件以及用户相似度,估计候选推荐对象和目标用户间的连接的属性值的步骤的流程示意图;
图9为一个实施例中推送推荐信息的装置的结构框图;
图10为另一个实施例中推送推荐信息的装置的结构框图。
具体实施方式
为了使本发明的目的、技术方案及优点更加清楚明白,以下结合附图及实施例,对本发明进行进一步详细说明。应当理解,此处所描述的具体实施例仅仅用以解释本发明,并不用于限定本发明。
为了便于理解本发明提供的推送推荐信息的方法,先对异质信息网络的 相关概念进行解释说明。参照图1,异质信息网络的底层数据结构是一张有向图,异质信息网络中包含了不同类型的对象和连接,而对象之间的连接代表了不同的关系。异质性与丰富的信息使得异质信息网络在很多场景里成为一个更好的数据表达形式。元路径是异质信息网络中一个独有的特性,它通过两个对象类型间的关系序列连接网络中的两个对象,被广泛地应用于探索丰富的语义信息。用户、用户属性、候选推荐对象以及候选推荐对象的属性都可以作为异质信息网络中的对象。连接表示相连接的对象间的关系。
异质信息网络可用于推荐电影,这种情况下异质信息网络包括了不同类型的对象,比如用户和电影;还包括了这些对象之间的各种关系,如观影信息、社交网络以及对象的属性信息。异质信息网络可以有效整合各种可能对推荐有用的信息。此外,探索网络中对象和关系的不同语义,可以揭示对象间微妙的关系。
例如,图1中的元路径“用户-电影-用户”表示看过相同电影的用户,这个元路径可以被用于基于观影记录的相似用户检索。如果根据这条元路径推荐电影,就会推荐那些被与目标用户有相同观影记录的其他用户看过的电影。类似地,基于元路径“用户-兴趣小组-用户”可以找到具有相似兴趣爱好的用户。因此,我们可以根据连接用户的不同元路径找到相似用户,然后直接推荐这些相似用户喜欢的推荐对象。进而,合理地设置不同的元路径可以实现不同的推荐模型。
然而,常规的异质信息网络和元路径并不考虑连接上的属性值,但用于推荐电影的异质信息网络可能包含带有属性值的连接。具体来说,用户会给他看过的电影打一个1至5之间的评分值(如图1中用户与电影间的连接所示),越高的评分值表示用户越喜欢这个电影。如果不考虑评分值大小,相似用户检索可能会出现不准确的结果。例如,基于路径“用户-电影-用户”,汤姆与玛丽或鲍勃的相似度是相同的,这是因为他们都观看了相同的电影。然而,他们可能会因为完全不同的兴趣而给出不一样的评分值。事实上,汤姆和鲍勃对相同的电影打了非常高的评分值,因此基于评分值来看,他们是相 似的。而玛丽则有完全不同的口味,这是因为她一点都不喜欢这些电影。常规的元路径并不考虑连接上的属性值,因此不能揭示这些微妙的差别。然而,这些差别是非常重要的,尤其是在候选推荐对象时。因此,我们需要扩展现有的异质信息网络和元路径等概念来引入连接上的属性值。
具体定义异质信息网络,给定一个网络模式
Figure PCTCN2016084284-appb-000001
其中包含一个对象类型集合
Figure PCTCN2016084284-appb-000002
一个连接对象对的关系集合
Figure PCTCN2016084284-appb-000003
以及关系上的属性值集合
Figure PCTCN2016084284-appb-000004
带属性值的信息网络是一个有向图G=(V,E,W),其中包含一个对象类型映射函数
Figure PCTCN2016084284-appb-000005
一个连接类型映射函数
Figure PCTCN2016084284-appb-000006
以及一个属性值类型映射函数
Figure PCTCN2016084284-appb-000007
每一个对象v∈V属于一个特定的对象类型
Figure PCTCN2016084284-appb-000008
每一个连接e∈E属于一个特定的关系
Figure PCTCN2016084284-appb-000009
每一个属性值w∈W属于一个特定的属性值类型
Figure PCTCN2016084284-appb-000010
当对象类型
Figure PCTCN2016084284-appb-000011
且关系类型
Figure PCTCN2016084284-appb-000012
时,可称为同质信息网络。当对象类型
Figure PCTCN2016084284-appb-000013
(或关系类型
Figure PCTCN2016084284-appb-000014
)且属性值类型
Figure PCTCN2016084284-appb-000015
时,网络可称为不带属性值的异质信息网络。当对象类型
Figure PCTCN2016084284-appb-000016
(或关系类型
Figure PCTCN2016084284-appb-000017
)且属性值类型
Figure PCTCN2016084284-appb-000018
时,网络可被称为带属性值的异质信息网络。
常规的异质信息网络是不带属性值的,即网络中的关系没有属性值或不考虑这些属性值。对于带属性值的异质信息网络来说,网络中的某些关系上带有属性值,这些属性值可能是连续的或是离散的。连续的属性值可以转换成离散的属性值处理。比如,在用于推荐电影的异质信息网络中,用户可以为看过的电影打一个1到5之间的评分值;在科学文献的异质信息网络中,作者与论文间的关系可以取不同的属性值来表示该作者是论文的第几作者。
参照图2,提供了一种用于推荐电影的带属性值的异质信息网络,图2展示了它的网络模式。这个异质信息网络包括了六种不同对象类型的对象(如用户、电影、兴趣小组、演员、导演以及电影类型)和它们之间的关系,对象间的连接表示不同的关系。例如,用户间的连接表示社交关系,用户和电影间的连接表示评分关系。此外,用户和电影间的评分关系上还带有一种取值范围为1至5间的整数的属性值。
异质信息网络中的两个对象可以通过不同的元路径连接,而这些路径有不同的意义。例如在图2中,用户之间可以通过“用户-用户(简称UU)”,“用户-群组-用户”(简称UGU),“用户-电影-用户”(简称UMU)等路径连接。这些元路径是对象类型间的一个关系序列。
接下来定义带属性值的异质网络中的元路径:元路径是指连接上带有属性值约束的元路径,可以表示成
Figure PCTCN2016084284-appb-000019
(也可简写成
Figure PCTCN2016084284-appb-000020
其中下标l表示元路径的编号。如果关系Re的连接上有属性值,那么属性值函数δ(Re)是关系Re上的属性值的集合,否则δ(Re)是一个空集。
Figure PCTCN2016084284-appb-000021
表示Al和Al+1间的关系Rel是基于属性值δl(Rel)的。属性值函数上的属性值约束条件
Figure PCTCN2016084284-appb-000022
是属性值函数间的一个关联约束集合。如果元路径上的所有属性值函数都是空集(相应的属性值约束条件
Figure PCTCN2016084284-appb-000023
也是空集),那么这样的元路径被称为不带属性值的元路径,否则就称为带属性值的元路径或者扩展元路径。
以图2为例,用户U和电影M之间的评分关系的评分值可取值1-5。元路径
Figure PCTCN2016084284-appb-000024
(即U(1)M)表示用户对电影评1分,其中隐含着用户不喜欢该电影的意思。元路径
Figure PCTCN2016084284-appb-000025
(即U(1,2)M(1,2)U)表示与目标用户不喜欢相同电影的候选用户,而不带属性值的元路径UMU则只可以反映有相同观影记录的用户。此外,通过灵活地设置关联约束,可以限制元路径中不同关系上的属性值。参照图1,例如,元路径U(i)M(j)U|i=j表示与目标用户对相同电影有完全相同评分值的用户,在这一条元路径下,可以容易地发现汤姆与鲍勃相似,但他们与玛丽不相似。
如图3所示,在一个实施例中,提供了一种推荐***,包括通过网络连接的服务器302和终端304。服务器302可以是独立的物理服务器,也可以是多个物理服务器构成的服务器集群。终端304可以是台式计算机或者移动终端,移动终端包括手机、智能手表、平板电脑以及PDA(个人数字助理)中的至少一种。
如图4所示,在一个实施例中,服务器302包括通过***总线连接的处 理器、非易失性存储介质、内存储器和网络接口。其中处理器具有计算功能和控制服务器302工作的功能,该处理器被配置为执行一种推送推荐信息的方法。非易失性存储介质包括磁存储介质、光存储介质和闪存式存储介质中的至少一种。非易失性存储介质存储有操作***和推送推荐信息的装置,该推送推荐信息的装置用于实现一种推送推荐信息的方法。网络接口用于连接到网络与终端304通信。
如图5所示,在一个实施例中,提供了一种推送推荐信息的方法,本实施例以该方法应用于上述图3和图4中的服务器302来举例说明。
步骤502,获取异质信息网络中连接候选用户和目标用户的元路径;该元路径包括候选用户和候选推荐对象间的具有属性值的连接。
具体地,异质信息网络中包括各种类型的对象,对象至少包括候选用户、目标用户和候选推荐对象,各对象间的连接表示相连接的对象之间的关系。这里的一个用户是自然人映射成为的一个数据对象,目标用户表示推荐信息的接收方,候选用户则是关系和属性值已知的用户。候选推荐对象是指可被推荐给目标用户的对象,包括电影、音乐、书籍、好友、群组以及商品中的至少一种。
异质信息网络包括元路径,该元路径连接候选用户和目标用户,且包括候选推荐对象。元路径的对象类型是对称的,比如,在如图2所示的异质信息网络中,元路径可以是“用户-群组-用户”、“用户-电影-用户”、“用户-电影-电影类型-电影-用户”、“用户-用户-用户”以及“用户-电影-导演-电影-用户”等中的至少一种。
步骤504,获取目标用户和候选用户相对于元路径的用户相似度。
相似度是指相似性度量,表示两个对象相似的程度。目标用户和候选用户相对于元路径的用户相似度,是基于该元路径所计算出的目标用户和候选用户间的相似度。由于太长的元路径没有意义且会产生不好的相似度,因此可以限定元路径的长度不超过4。元路径的长度与元路径中连接的数量相等。
在一个实施例中,步骤504包括:根据目标用户和候选用户间相对于元 路径的各条原子元路径的相似度,以获取目标用户和候选用户相对于元路径的用户相似度。
其中,若元路径
Figure PCTCN2016084284-appb-000026
中的属性值函数δl(Rel)取一个固定的值,那么这样的路径就被称为原子元路径(Atomic meta path)。一个元路径是所有满足该元路径的属性值约束条件的原子元路径的集合。对于一条原子元路径,现有的相似性度量方法就可以被直接使用。现有的相似性度量方法比如PathSim(Y.Sun,J.Han,X.Yan,P.Yu,and T.Wu.Pathsim:Meta path-based top-k similarity search in heterogeneous information networks.In VLDB,pages 992-1003,2011)、PCRW(N.Lao and W.Cohen.Fast query execution for retrieval models based on path constrained random walks.In KDD,pages 881-888,2010)和HeteSim(C.Shi,X.Kong,Y.Huang,P.S.Yu,and B.Wu.Hetesim:A general framework for relevance measure in heterogeneous networks.IEEE Transactions on Knowledge and Data Engineering,26(10):2479-2492,2014)。
步骤506,根据候选用户和候选推荐对象间的连接的属性值、元路径的属性值约束条件以及用户相似度,估计候选推荐对象和目标用户间的连接的属性值。
具体地,元路径还包括属性值约束条件,用于约束该元路径上候选用户和候选推荐对象间的属性值,以及,候选推荐对象和目标用户间的连接的属性值这两种属性值之间的数学关系。该属性值约束条件可以是该两种属性值相等或者相差在预设范围内。其中两种属性值相等,是严格的属性值约束,使得推荐结果更加准确;而两种属性值相差在预设范围内,则可以挖掘出更宽泛的元路径语义。
以元路径U(i)M(j)U|i=j为例,属性值i和属性值j是取值为1-5的变量,且属性值i和属性值j必须满足属性值约束条件:i=j。或者,以元路径U(i)M(j)U||i-j|≤1)为例,属性值i和属性值j必须满足属性值约束条件:|i-j|≤1。
这样用户相似度可以反映出候选用户和目标用户的相似程度,从而可用来确定相应候选用户对应的属性值在估计候选推荐对象和目标用户间的连接的属性值时考虑的程度。属性值约束条件可用来限定估计的属性值的具体值,使得估计的属性值符合该元路径所表达的语义。
步骤508,当估计的属性值满足推荐条件时,向目标用户对应的终端发送候选推荐对象的推荐信息。
具体地,推荐条件是是否向目标用户推荐相应的候选推荐对象的判断条件,推荐条件比如可以是估计的属性值大于预设阈值,或者估计的属性值等于预设阈值或者估计的属性值小于预设阈值,具体根据属性值的含义和推荐需求来决定。
比如,若属性值表示用户对电影的评分值,如果评分值是与用户对电影的态度正相关的,即用户越喜欢电影评分值越高,那么就会在当估计的属性值大于等于预设阈值时,向目标用户对应的终端发送候选推荐对象的推荐信息。如果评分值是与用户对电影的态度负相关的,即用户越喜欢电影评分值越低,那么就会在当估计的属性值小于预设阈值时,向目标用户对应的终端发送候选推荐对象的推荐信息。预设阈值可以根据推荐精度要求灵活配置。
推荐信息可以包括候选推荐对象的描述信息,还可以包括候选推荐对象的访问地址。比如当候选推荐对象为电影时,描述信息可以包括电影名称、电影概述、导演、演员以及宣传海报等信息,访问地址则可以是购票网站的访问地址或者在线视频网站的访问地址。
上述推送推荐信息的方法,通过一种新型的异质信息网络来实现对象的推荐,该异质信息网络包括连接候选用户和目标用户的元路径,可以表示出目标用户和候选用户间的社交关系。该元路径中候选用户和候选推荐对象间的连接具有属性值,以对候选用户和候选推荐对象间的关系进行量化。在获取到目标用户和候选用户相对于元路径的用户相似度之后,就可以根据该用户相似度结合候选用户和候选推荐对象间的连接的属性值以及属性值约束条件来估计候选推荐对象和目标用户间的连接的属性值,估计出的属性值可反 映出目标用户和候选推荐对象间经过量化的关系。这样在利用异质信息网络来进行推荐时,不仅考虑了目标用户的社交关系,还考虑了目标用户与候选推荐对象之间量化的关系,从而使得推送结果更加准确。
如图6所示,在一个实施例中,步骤504具体包括以下步骤:
步骤602,根据元路径的属性值约束条件将元路径拆分为多条原子元路径。
具体地,对于一条元路径
Figure PCTCN2016084284-appb-000027
可在满足属性值约束条件
Figure PCTCN2016084284-appb-000028
的条件下遍历属性值的离散取值范围,从而拆分出多条原子元路径。拆分出的原子元路径的数量与离散取值范围及属性值约束条件
Figure PCTCN2016084284-appb-000029
相关。
举例说明,以图2所示的异质信息网络为例,U(1)M(1)U和U(1)M(2)U都是原子元路径。元路径U(i)M(j)U|i=j可以看成是5条原子元路径的集合,可以拆分出5个原子元路径,即U(1)M(1)U、U(2)M(2)U、U(3)M(3)U、U(4)M(4)U和U(5)M(5)U。
步骤604,获取目标用户和候选用户间相对于各条原子元路径的相似度。
具体地,可以采用PathSim、PCRW以及HeteSim这些相似性度量方法中的任意一种来计算目标用户和候选用户间相对于各条原子元路径的相似度。其中采用PathSim计算相似度时,具体先沿着原子元路径计算连接目标用户和候选用户的路径实例数量,然后对这个数量进行规则化,以获得相应的相似度。
步骤606,根据获取的相对于各条原子元路径的相似度计算目标用户和候选用户相对于元路径的用户相似度。
相对于各条原子元路径的相似度也就是基于各条原子元路径的相似度。由于元路径可以拆分为一组相应的原子元路径,基于元路径的用户相似度可以看成是基于对应的所有原子元路径的相似度的综合相似度。用户相似度可以采用求和或者求加权和的方式获得。
在一个实施例中,步骤606包括:计算获取的相对于各条原子元路径的 相似度的和;将相似度的和直接或者进行正相关运算后作为目标用户和候选用户相对于元路径的用户相似度。
具体地,在计算出目标用户和候选用户相对于各条原子元路径的相似度后,将元路径的所有的原子元路径相对应的相似度求和。然后可将这些相似度的和直接作为目标用户和候选用户相对于元路径的用户相似度,也可以对这些相似度的和进行正相关运算后作为目标用户和候选用户相对于元路径的用户相似度。其中正相关运算是指因变量与自变量的变化趋势一致的运算,比如加上或减去或乘以或除以一个正值。正相关运算包括规则化处理,采用PathSim和HeteSim这两种相似性度量方法计算出的用户相似度需要对用户相似度进行规则化处理,以限定计算出的相似度的取值范围。
这里以用于推荐电影的异质信息网络来举例说明计算用户相似度的过程。参照图7,用户u1、u2和u3都观看了电影m1和电影m2,并给出了相应的评分值,如图7中的评分值矩阵。参照图7上半部分,在传统的不带属性值的元路径UMU中,不考虑这些评分值,只考虑是否观看了相应的电影,这样采用PathSim计算出的相似度矩阵表示u1、u2和u3两两间的相似度都是相等的。
参照图7下半部分,相比之下,带属性值的元路径U(i)M(j)U|i=j被拆分为5条原子元路径,每条原子元路径上的属性值是固定的。这样就可以根据该原子元路径上u1、u2和u3是否分别对电影m1和电影m2有评分值,从而直接采用PathSim计算出每条原子元路径上u1、u2和u3两两间的相似度。然后将不同原子元路径下u1、u2和u3两两间的相似度分别求和并进行规则化,就可以获得元路径U(i)M(j)U|i=j下的相似度矩阵。可以看出只有用户u1和用户u2是相似的,因为用户u1和用户u2对电影有相同的观影偏好。
如图8所示,在一个实施例中,步骤506具体包括以下步骤:
步骤802,获取目标用户和候选推荐对象间的连接的属性值的离散取值范围。
具体地,假设用户集合U,候选用户v和目标用户u均属于用户集合U; 候选推荐对象的集合|X|,候选推荐对象x∈|X|;
Figure PCTCN2016084284-appb-000030
表示元路径的集合;R∈R|U|×|X|是属性值矩阵,Ru,x∈R表示目标用户u与候选推荐对象x的属性值;属性值Ru,x的离散取值范围为1到N的正整数,比如N可以取5。R表示实数集。
步骤804,对于离散取值范围内的每个取值,分别获取具有与取值满足属性值约束条件的属性值的候选用户和候选推荐对象间的连接,根据获取的连接所对应的候选用户与目标用户间的用户相似度计算取值对应的属性值强度。
具体地,假设S∈R|U|×|U|是用户相似度矩阵,表示用户集合U中两两用户之间的相似度,其中
Figure PCTCN2016084284-appb-000031
表示目标用户u和候选用户v相对于元路径
Figure PCTCN2016084284-appb-000032
的用户相似度。这里定义属性值强度Q∈R|U|×|X|×N,其中
Figure PCTCN2016084284-appb-000033
表示在给定路径
Figure PCTCN2016084284-appb-000034
下目标用户u与候选推荐对象x的连接的属性值为r的强度。属性值强度
Figure PCTCN2016084284-appb-000035
与用户相似度
Figure PCTCN2016084284-appb-000036
相关;属性值强度
Figure PCTCN2016084284-appb-000037
还与具有满足属性值约束条件的属性值的候选用户v的数量相关,当属性值约束条件为两种属性值相等时,此时属性值强度
Figure PCTCN2016084284-appb-000038
与属性值为r的候选用户v的数量相关。
Figure PCTCN2016084284-appb-000039
可采用下述公式(1)计算:
Figure PCTCN2016084284-appb-000040
公式(1)中Ev,x,r表示候选用户v是否与候选推荐对象x的连接的属性值为r,若是则Ev,x,r为1,否则为0。此处仅以属性值约束条件为两种属性值相等时进行举例,当属性值约束条件为其它情况时可相应修改Ev,x,r为仅在Rv,x与r满足属性值约束条件时值为1。
这样对应于离散取值范围内的每个取值r,分别获取元路径
Figure PCTCN2016084284-appb-000041
中具有与取值r满足属性值约束条件
Figure PCTCN2016084284-appb-000042
的属性值的连接所连接的候选用户v与目标用户u间的用户相似度
Figure PCTCN2016084284-appb-000043
并求和,获得与取值r对应的属性值强度
Figure PCTCN2016084284-appb-000044
步骤806,将离散取值范围内的各个取值分别以相应的属性值强度为权重计算加权平均值。
具体地,将1到N的各个取值r分别乘以相应的属性值强度
Figure PCTCN2016084284-appb-000045
进行加 权来计算加权平均值。在一个实施例中,还可以对属性值强度
Figure PCTCN2016084284-appb-000046
进行规则化处理后,将离散取值范围内的各个取值分别以相应的规则化的属性值强度为权重计算加权平均值。对属性值强度
Figure PCTCN2016084284-appb-000047
进行规则化处理具体为:用属性值强度
Figure PCTCN2016084284-appb-000048
除以相应元路径下所有属性值强度的和。
步骤808,根据计算出的加权平均值获得候选推荐对象和目标用户间的连接的估计的属性值。
具体地,当仅存在一条元路径时,步骤806计算出的加权平均值可以直接作为估计的属性值。具体估计的属性值可采用以下公式(2)来计算:
Figure PCTCN2016084284-appb-000049
其中,
Figure PCTCN2016084284-appb-000050
表示在元路径
Figure PCTCN2016084284-appb-000051
下候选推荐对象x和目标用户u间的连接的估计的属性值,N表示离散取值范围的上限,
Figure PCTCN2016084284-appb-000052
表示与取值r对应的属性值强度,
Figure PCTCN2016084284-appb-000053
表示与取值k对应的属性值强度。
本实施例中,可以实现根据给定的一条元路径来预测目标用户和候选推荐对象间连接的属性值,从而向目标用户推荐满足推荐条件的候选推荐对象。而且,公式(2)有一个额外的优势,即它可以消除不同元路径下计算得到的用户相似度的偏倚。考虑到基于不同元路径计算得到的用户相似度有不同的取值范围,这样会使得不同元路径间的相似度计算和属性值强度难以比较,公式(2)中的规则化属性值强度可以消除取值范围的差异性。
在一个实施例中,步骤808包括:将各个元路径下计算出的加权平均值分别乘以相应的元路径的路径权重以计算加权平均值,获得候选推荐对象和目标用户间的连接的估计的属性值。
具体地,为所有用户对每一条元路径设置一个统一的路径权重,表示用户对这条元路径的偏好。具体如公式(3):
Figure PCTCN2016084284-appb-000054
其中,其中w(l)表示元路径
Figure PCTCN2016084284-appb-000055
的路径权重。基于所有元路径的综合的估计 的属性值
Figure PCTCN2016084284-appb-000056
可以用每一条元路径上的估计的属性值
Figure PCTCN2016084284-appb-000057
的加权平均值来表示。经过目标优化后获得的各条元路径的路径权重的和为1,因此
Figure PCTCN2016084284-appb-000058
就是将各个元路径下计算出的加权平均值分别乘以相应的元路径的路径权重以计算加权平均值。
为了使得估计的属性值矩阵
Figure PCTCN2016084284-appb-000059
接近于真实的属性值矩阵R,这里基于真实的属性值和估计的属性值间的平方误差定义了一个目标函数,如公式(4):
Figure PCTCN2016084284-appb-000060
其中,符号⊙表示矩阵的阿达马乘积,即对应元素的乘积;||·||p表示矩阵的p范数。Y表示一个指示矩阵,Yu,x=1表示目标用户u与候选推荐对象x的连接有属性值,否则Yu,x=0;λ0是控制参数。s.t.表示约束条件。在已知目标用户和候选推荐对象的连接的真实的属性值的情况下,可以通过优化上述公式(4)的目标函数来求解路径权重向量
Figure PCTCN2016084284-appb-000061
在一个实施例中,步骤808包括:将各个元路径下计算出的加权平均值分别乘以与目标用户和相应的元路径对应的路径权重以计算加权平均值,获得候选推荐对象和目标用户间的连接的估计的属性值。
具体地,考虑到在很多现实的应用场景中,每个用户都有自己的个性化的兴趣偏好,统一的路径权重不能为用户提供个性化推荐。为了实现个性化推荐,可以为每个用户设置路径权重向量。假设路径权重矩阵表示为
Figure PCTCN2016084284-appb-000062
其中每个元素
Figure PCTCN2016084284-appb-000063
表示与目标用户u和路径
Figure PCTCN2016084284-appb-000064
对应的路径权重。列向量W(l)∈R|U|×1表示所有用户在路径
Figure PCTCN2016084284-appb-000065
下的路径权重向量。因此估计的属性值矩阵
Figure PCTCN2016084284-appb-000066
表示目标用户u在综合所有元路径下与候选推荐对象v间的连接的属性值。存在公式(5):
Figure PCTCN2016084284-appb-000067
再定义一个目标函数,如公式(6):
Figure PCTCN2016084284-appb-000068
其中,符号⊙表示矩阵的阿达马乘积,即对应元素的乘积;||·||p表示矩阵的p范数。Y表示一个指示矩阵,Yu,x=1表示目标用户u与候选推荐对象x的连接有属性值,否则Yu,x=0;λ0是控制参数;diag(W(l))表示由向量W(l)转化成的对角矩阵。s.t.表示约束条件。在已知目标用户u和候选推荐对象x的连接的真实的属性值的情况下,可以通过优化上述公式(6)的目标函数来求解路径权重矩阵W。
在一个实施例中,该推送推荐信息的方法还包括:获取候选推荐对象和目标用户间的连接的真实的属性值;将与目标用户和元路径对应的路径权重初始化;根据用户相似度,朝趋近于与候选用户和元路径对应的路径权重的平均值的方向调整初始化的路径权重,使得真实的属性值和估计的属性值的差距满足最小化条件。
具体地,尽管公式(6)考虑到了用户个性化的路径权重,但是对那些只有少量属性值信息的用户很难进行有效的权重学习。需要学习的权重一共有
Figure PCTCN2016084284-appb-000069
而训练的样本数则远小于|U|×|X|。训练样本经常不足以进行权重学习,这对于那些冷启动用户和物品来说尤其严重。用户的路径权重与其相似用户的路径权重应该比较一致。对于那些只有少量属性值的用户来说,他们的路径权重可以从那些与他们相似的其他用户的路径权重中学习得到,这是因为基于元路径的用户相似度对于这些用户来说更加有效。
将与目标用户和元路径对应的路径权重初始化,具体可初始化为0或者大于0的值,然后朝趋近于与候选用户和元路径对应的路径权重的平均值的方向调整初始化的路径权重,趋近的速度与用户相似度正相关,用户相似度越大则趋近越快,用户相似度越低则趋近越慢。当真实的属性值和估计的属 性值的差距满足最小化条件时停止调整。最小化条件可以采用上述公式(4)或者(6)或者下述的公式(9)。
因此,定义路径权重正则化项,如公式(7):
Figure PCTCN2016084284-appb-000070
其中|U|表示用户总数,
Figure PCTCN2016084284-appb-000071
表示扩展路径总数,
Figure PCTCN2016084284-appb-000072
表示与目标用户u和路径
Figure PCTCN2016084284-appb-000073
对应的路径权重,
Figure PCTCN2016084284-appb-000074
表示与候选用户v和路径
Figure PCTCN2016084284-appb-000075
对应的路径权重,
Figure PCTCN2016084284-appb-000076
是基于路径
Figure PCTCN2016084284-appb-000077
且经过规则化后的目标用户u和候选用户v的用户相似度。为了方便,路径权重正则化项可以用以下公式(8)的矩阵形式表示:
Figure PCTCN2016084284-appb-000078
其中,W(l)是路径权重矩阵,
Figure PCTCN2016084284-appb-000079
是基于路径
Figure PCTCN2016084284-appb-000080
且经过规则化后的用户相似度矩阵,||·||2表示矩阵的2范数。
在上述公式(6)的基础上,增加路径权重正则化项,得到如以下公式(9)的目标函数:
Figure PCTCN2016084284-appb-000081
其中符号⊙表示矩阵的阿达马乘积,即对应元素的乘积;||·||p表示矩阵的p范数;Y表示一个指示矩阵,Yu,x=1表示目标用户u与候选推荐对象x的连接有属性值,否则Yu,x=0;λ0是控制参数;λ1是另一个控制参数;W(l)表示所有用户在路径
Figure PCTCN2016084284-appb-000082
下的路径权重向量;diag(W(l))表示由向量W(l)转化成的对角矩阵;W表示所有用户的路径权重矩阵;
Figure PCTCN2016084284-appb-000083
表示估计的基于元路径
Figure PCTCN2016084284-appb-000084
的属性值矩阵,R表示真实的属性值矩阵。
上述公式(9)是一个非负的二次规划问题,即非负矩阵分解的一种简单 形式,可以使用解决带非负界限约束优化问题的梯度投影法来进行优化求解,解决带非负界限约束优化问题的梯度投影法可参考“C.J.Lin.Projected gradient methods for non-negative matrix factorization.In Neural Computation,pages 2756-2279,2007”。公式(9)对
Figure PCTCN2016084284-appb-000085
的梯度为:
Figure PCTCN2016084284-appb-000086
其中符号T表示转置。的更新公式如公式(11):
Figure PCTCN2016084284-appb-000088
其中α是步长,可以根据需要设置。
具体可通过以下步骤(1)至步骤(7)来学习与目标用户和元路径对应的路径权重。步骤(1)至步骤(7)可称为SemRec方法(Semantic path based personalized Recommendation method,基于语义路径的个性化推荐方法)。
步骤(1),获取带属性值的异质信息网络G、连接用户的元路径集合
Figure PCTCN2016084284-appb-000089
控制参数λ0、控制参数λ1、参数更新时的步长α以及收敛阈值ε。
步骤(2),相对于元路径集合
Figure PCTCN2016084284-appb-000090
中的每条元路径,分别计算用户相似度矩阵S(l)、属性值强度矩阵Q(l)和估计的属性值矩阵
Figure PCTCN2016084284-appb-000091
步骤(3),初始化路径权重矩阵W>0。
重复执行以下步骤(4)、(5)和(6),直至满足|W-Wold|<ε。
步骤(4),Wold:=W。
步骤(5),计算
Figure PCTCN2016084284-appb-000092
步骤(6),
Figure PCTCN2016084284-appb-000093
步骤(7),输出所有用户的路径权重矩阵W。
其中Wold:=W表示将W赋值给Wold
Figure PCTCN2016084284-appb-000094
表示求取公式(9)的偏微分,
Figure PCTCN2016084284-appb-000095
表示取0和
Figure PCTCN2016084284-appb-000096
之间的最大值,|W-Wold|<ε表示相邻两次迭代计算的W相差小于收敛阈值ε。
从目标函数中可以发现,统一的路径权重学***,小的λ1表示更强烈的用户个性化路径权重,但是这样会使得权重学习变得非常困难。因此现实的应用需要根据应用场景设置一个合适的λ1
在一个实施例中,候选推荐对象为网络资源;属性值为评分值。其中网络资源包括电影、音频以及小说等可从网络获取的资源,评分值可用来反映用户对网络资源的量化的态度。
为了验证带属性值的异质信息网络的推荐效果,从网络上爬取了两个数据集。第一个数据集包含了13367个用户对象用户、12677部电影以及1068278个1-5的评分值。第一个数据集还包含了用户对象用户的社交关系以及用户对象用户和电影的属性信息。第二个数据集包含用户对象用户对本地商户的评分值,以及用户对象用户和商户的属性信息。这个数据集包含16239个用户对象用户、14284个本地商户以及198397个1-5的评分值。表1是两个数据集的详细统计信息。两个数据集有一些不一样的性质,第一个数据集的评分值关系更密集但社交关系非常稀疏,而第二个数据集的评分值关系比较稀疏但社交关系更密集。
表1:
Figure PCTCN2016084284-appb-000103
Figure PCTCN2016084284-appb-000104
这里使用了两种一般的评估指标,平均方根误差(RMSE)和平均绝对误差(MAE)来评估评估计属性值的质量。
Figure PCTCN2016084284-appb-000105
Figure PCTCN2016084284-appb-000106
其中Rtest表示整个测试集,这里将一个数据集分成训练集和测试集,训练集用来训练带属性值的异质信息网络,测试集用来测试训练过的异质信息网络的效果。较小的MAE或RMSE表示更好的效果。
为了展示提出的SemRec方法的有效性,比较了SemRec的四种变种方法。除了带路径权重规则化项的个性化路径权重学习方法(称为SemRecReg)以外,我们还考虑了三个SemRec的特殊版本:基于单条元路径的方法(称为 SemRecSgl),对所有用户学习统一路径权重的方法(称为SemRecAll),以及对每个用户学习个性化路径权重的方法(称为SemRecInd)。
由于太长的元路径没有意义且会产生不好的相似度,因此这里对每个数据集都使用了5条长度不超过4的元路径。表2展示了这些带权或不带权的元路径。在SemRec中使用PathSim来计算用户相似度。SemRec中的参数λ0设置成0.01,λ1设置成103
表2:
第一个数据集 第二个数据集
UGU UU
U(i)M(j)U|i=j UCoU
U(i)MDM(j)U|i=j U(i)B(j)U|i=j
U(i)MAM(j)U|i=j U(i)BCaB(j)U|i=j
U(i)MTM(j)U|i=j U(i)BCiB(j)U|i=j
对于第一个数据集,设置了不同的训练数据比例(20%,40%,60%,80%)来展示不同数据稀疏度下的对比结果。训练数据比例为20%表示用户-候选推荐对象评分值矩阵中有20%的评分值作为训练集进行模型训练,预测剩下的80%的评分值。第一个数据集有更密集的评分值关系。而第二个数据集的评分值关系更稀疏,因此在第二个数据集上使用更多的数据作训练集(60%,70%,80%,90%)。对每个实验结果,按给定的比例独立随机的划分了10次训练集和测试集,并以平均值作为表3所示的结果。
表3:
Figure PCTCN2016084284-appb-000107
Figure PCTCN2016084284-appb-000108
通过分析表3的测试结果,不同版本的SemRec有不同的性能表现。一般情况下,多条路径的SemRec(如SemRecAll和SemRecReg)比单一路径的 SemRec(即SemRecSgl)有更好的效果,但SemRecInd除外,这表示SemRec的路径权重学习方法可以有效地整合不同路径上产生的相似度信息。SemRecInd由于评分值的稀疏性,其推荐效果在大部分情况下比SemRecAll的效果更差。此外,SemRecReg在所有情况下都可以达到最好的效果。这是因为SemRecReg不仅实现了所有用户的个性化路径权重学习,还使用路径权重正则化来避免了评分值稀疏性带来的问题。
此外,记录了学***均运行时间。四个版本的SemRec随着路径权重学***衡效率和性能。
如图9所示,在一个实施例中,提供了一种推送推荐信息的装置900,具有实现上述各个实施例的推送推荐信息的方法的功能模块。该推送推荐信息的装置900包括:元路径获取模块901、用户相似度获取模块902、属性值估计模块903和推送模块904。
元路径获取模块901,用于获取异质信息网络中连接候选用户和目标用户的元路径;元路径包括候选用户和候选推荐对象间的具有属性值的连接。
用户相似度获取模块902,用于获取目标用户和候选用户相对于元路径的用户相似度。
属性值估计模块903,用于根据候选用户和候选推荐对象间的连接的属性值、元路径的属性值约束条件以及用户相似度,估计候选推荐对象和目标用户间的连接的属性值。
推送模块904,用于当估计的属性值满足推荐条件时,向目标用户对应的终端发送候选推荐对象的推荐信息。
如图10所示,在一个实施例中,用户相似度获取模块902包括:拆分模 块902a、相似度计算模块902b和用户相似度合成模块902c。
拆分模块902a,用于根据元路径的属性值约束条件将元路径拆分为多条原子元路径。
相似度计算模块902b,用于获取目标用户和候选用户间相对于各条原子元路径的相似度。
用户相似度合成模块902c,用于根据获取的相对于各条原子元路径的相似度计算目标用户和候选用户相对于元路径的用户相似度。
在一个实施例中,用户相似度合成模块902c还用于计算获取的相对于各条原子元路径的相似度的和;将相似度的和直接或者进行正相关运算后作为目标用户和候选用户相对于元路径的用户相似度。
在一个实施例中,属性值估计模块903包括:离散取值范围获取模块903a、属性值强度计算模块903b、加权平均模块903c和估计结果生成模块903d。
离散取值范围获取模块903a,用于获取目标用户和候选推荐对象间的连接的属性值的离散取值范围。
属性值强度计算模块903b,用于对于离散取值范围内的每个取值,分别获取具有与取值满足属性值约束条件的属性值的候选用户和候选推荐对象间的连接,根据获取的连接所对应的候选用户与目标用户间的用户相似度计算取值对应的属性值强度。
加权平均模块903c,用于将离散取值范围内的各个取值分别以相应的属性值强度为权重计算加权平均值。
估计结果生成模块903d,用于根据计算出的加权平均值获得候选推荐对象和目标用户间的连接的估计的属性值。
在一个实施例中,估计结果生成模块903d还用于将各个元路径下计算出的加权平均值分别乘以相应的元路径的路径权重以计算加权平均值,获得候选推荐对象和目标用户间的连接的估计的属性值。
在一个实施例中,估计结果生成模块903d还用于将各个元路径下计算出 的加权平均值分别乘以与目标用户和相应的元路径对应的路径权重以计算加权平均值,获得候选推荐对象和目标用户间的连接的估计的属性值。
在一个实施例中,推送推荐信息的装置900还包括:路径权重学***均值的方向调整初始化的路径权重,使得真实的属性值和估计的属性值的差距满足最小化条件。
在一个实施例中,候选推荐对象为网络资源;属性值为评分值。
本领域普通技术人员可以理解实现上述实施例方法中的全部或部分流程,是可以通过计算机程序来指令相关的硬件来完成,所述的程序可存储于一计算机可读取存储介质中,该程序在执行时,可包括如上述各方法的实施例的流程。其中,所述的存储介质可为磁碟、光盘、只读存储记忆体(Read-Only Memory,ROM)等非易失性存储介质,或随机存储记忆体(Random Access Memory,RAM)等。
以上所述实施例的各技术特征可以进行任意的组合,为使描述简洁,未对上述实施例中的各个技术特征所有可能的组合都进行描述,然而,只要这些技术特征的组合不存在矛盾,都应当认为是本说明书记载的范围。
以上所述实施例仅表达了本发明的几种实施方式,其描述较为具体和详细,但并不能因此而理解为对发明专利范围的限制。应当指出的是,对于本领域的普通技术人员来说,在不脱离本发明构思的前提下,还可以做出若干变形和改进,这些都属于本发明的保护范围。因此,本发明专利的保护范围应以所附权利要求为准。

Claims (24)

  1. 一种推送推荐信息的方法,包括:
    获取异质信息网络中连接候选用户和目标用户的元路径;所述元路径包括所述候选用户和候选推荐对象间的具有属性值的连接;
    获取所述目标用户和所述候选用户相对于所述元路径的用户相似度;
    根据所述候选用户和候选推荐对象间的连接的属性值、所述元路径的属性值约束条件以及所述用户相似度,估计所述候选推荐对象和所述目标用户间的连接的属性值;及
    当估计的属性值满足推荐条件时,向所述目标用户对应的终端发送所述候选推荐对象的推荐信息。
  2. 根据权利要求1所述的方法,其特征在于,所述获取所述目标用户和所述候选用户相对于所述元路径的用户相似度包括:
    根据所述元路径的属性值约束条件将所述元路径拆分为多条原子元路径;
    获取所述目标用户和所述候选用户间相对于各条原子元路径的相似度;及
    根据获取的相对于各条原子元路径的相似度计算所述目标用户和所述候选用户相对于所述元路径的用户相似度。
  3. 根据权利要求2所述的方法,其特征在于,所述根据获取的相对于各条原子元路径的相似度计算所述目标用户和所述候选用户相对于所述元路径的用户相似度包括:
    计算获取的相对于各条原子元路径的相似度的和;及,
    将所述相似度的和直接或者进行正相关运算后作为所述目标用户和所述候选用户相对于所述元路径的用户相似度。
  4. 根据权利要求1所述的方法,其特征在于,所述根据所述候选用户和候选推荐对象间的连接的属性值、所述元路径的属性值约束条件以及所述用户相似度,估计所述候选推荐对象和所述目标用户间的连接的属性值包括:
    获取所述目标用户和候选推荐对象间的连接的属性值的离散取值范围;
    对于所述离散取值范围内的每个取值,分别获取具有与所述取值满足所述属性值约束条件的属性值的所述候选用户和候选推荐对象间的连接,根据获取的连接所对应的候选用户与所述目标用户间的用户相似度计算所述取值对应的属性值强度;
    将所述离散取值范围内的各个取值分别以相应的属性值强度为权重计算加权平均值;及
    根据计算出的加权平均值获得所述候选推荐对象和所述目标用户间的连接的估计的属性值。
  5. 根据权利要求4所述的方法,其特征在于,所述根据计算出的加权平均值获得所述候选推荐对象和所述目标用户间的连接的估计的属性值包括:
    将各个元路径下计算出的加权平均值分别乘以相应的元路径的路径权重以计算加权平均值,获得所述候选推荐对象和所述目标用户间的连接的估计的属性值。
  6. 根据权利要求4所述的方法,其特征在于,所述根据计算出的加权平均值获得所述候选推荐对象和所述目标用户间的连接的估计的属性值包括:
    将各个元路径下计算出的加权平均值分别乘以与目标用户和相应的元路径对应的路径权重以计算加权平均值,获得所述候选推荐对象和所述目标用户间的连接的估计的属性值。
  7. 根据权利要求6所述的方法,其特征在于,还包括:
    获取所述候选推荐对象和所述目标用户间的连接的真实的属性值;
    将与目标用户和元路径对应的路径权重初始化;及
    根据所述用户相似度,朝趋近于与所述候选用户和所述元路径对应的路径权重的平均值的方向调整初始化的路径权重,使得真实的属性值和估计的属性值的差距满足最小化条件。
  8. 根据权利要求1所述的方法,其特征在于,所述候选推荐对象为网络资源;所述属性值为评分值。
  9. 一种服务器,包括存储器和处理器,所述存储器中储存有指令,其特征在于,所述指令被所述处理器执行时,使得所述处理器执行以下步骤:
    获取异质信息网络中连接候选用户和目标用户的元路径;所述元路径包括所述候选用户和候选推荐对象间的具有属性值的连接;
    获取所述目标用户和所述候选用户相对于所述元路径的用户相似度;
    根据所述候选用户和候选推荐对象间的连接的属性值、所述元路径的属性值约束条件以及所述用户相似度,估计所述候选推荐对象和所述目标用户间的连接的属性值;及
    当估计的属性值满足推荐条件时,向所述目标用户对应的终端发送所述候选推荐对象的推荐信息。
  10. 根据权利要求9所述的服务器,其特征在于,所述获取所述目标用户和所述候选用户相对于所述元路径的用户相似度包括:
    根据所述元路径的属性值约束条件将所述元路径拆分为多条原子元路径;
    获取所述目标用户和所述候选用户间相对于各条原子元路径的相似度;及
    根据获取的相对于各条原子元路径的相似度计算所述目标用户和所述候选用户相对于所述元路径的用户相似度。
  11. 根据权利要求10所述的服务器,其特征在于,所述根据获取的相对于各条原子元路径的相似度计算所述目标用户和所述候选用户相对于所述元路径的用户相似度包括:
    计算获取的相对于各条原子元路径的相似度的和;及,
    将所述相似度的和直接或者进行正相关运算后作为所述目标用户和所述候选用户相对于所述元路径的用户相似度。
  12. 根据权利要求9所述的服务器,其特征在于,所述根据所述候选用户和候选推荐对象间的连接的属性值、所述元路径的属性值约束条件以及所述用户相似度,估计所述候选推荐对象和所述目标用户间的连接的属性值包 括:
    获取所述目标用户和候选推荐对象间的连接的属性值的离散取值范围;
    对于所述离散取值范围内的每个取值,分别获取具有与所述取值满足所述属性值约束条件的属性值的所述候选用户和候选推荐对象间的连接,根据获取的连接所对应的候选用户与所述目标用户间的用户相似度计算所述取值对应的属性值强度;
    将所述离散取值范围内的各个取值分别以相应的属性值强度为权重计算加权平均值;及
    根据计算出的加权平均值获得所述候选推荐对象和所述目标用户间的连接的估计的属性值。
  13. 根据权利要求12所述的服务器,其特征在于,所述根据计算出的加权平均值获得所述候选推荐对象和所述目标用户间的连接的估计的属性值包括:
    将各个元路径下计算出的加权平均值分别乘以相应的元路径的路径权重以计算加权平均值,获得所述候选推荐对象和所述目标用户间的连接的估计的属性值。
  14. 根据权利要求12所述的服务器,其特征在于,所述根据计算出的加权平均值获得所述候选推荐对象和所述目标用户间的连接的估计的属性值包括:
    将各个元路径下计算出的加权平均值分别乘以与目标用户和相应的元路径对应的路径权重以计算加权平均值,获得所述候选推荐对象和所述目标用户间的连接的估计的属性值。
  15. 根据权利要求14所述的服务器,其特征在于,还包括:
    获取所述候选推荐对象和所述目标用户间的连接的真实的属性值;
    将与目标用户和元路径对应的路径权重初始化;及
    根据所述用户相似度,朝趋近于与所述候选用户和所述元路径对应的路径权重的平均值的方向调整初始化的路径权重,使得真实的属性值和估计的 属性值的差距满足最小化条件。
  16. 根据权利要求9所述的服务器,其特征在于,所述候选推荐对象为网络资源;所述属性值为评分值。
  17. 一个或多个存储有计算机可执行指令的非易失性可读存储介质,所述计算机可执行指令被一个或多个处理器执行时,使得所述一个或多个处理器执行以下步骤:
    获取异质信息网络中连接候选用户和目标用户的元路径;所述元路径包括所述候选用户和候选推荐对象间的具有属性值的连接;
    获取所述目标用户和所述候选用户相对于所述元路径的用户相似度;
    根据所述候选用户和候选推荐对象间的连接的属性值、所述元路径的属性值约束条件以及所述用户相似度,估计所述候选推荐对象和所述目标用户间的连接的属性值;及
    当估计的属性值满足推荐条件时,向所述目标用户对应的终端发送所述候选推荐对象的推荐信息。
  18. 根据权利要求17所述的非易失性可读存储介质,其特征在于,所述获取所述目标用户和所述候选用户相对于所述元路径的用户相似度包括:
    根据所述元路径的属性值约束条件将所述元路径拆分为多条原子元路径;
    获取所述目标用户和所述候选用户间相对于各条原子元路径的相似度;及
    根据获取的相对于各条原子元路径的相似度计算所述目标用户和所述候选用户相对于所述元路径的用户相似度。
  19. 根据权利要求18所述的非易失性可读存储介质,其特征在于,所述根据获取的相对于各条原子元路径的相似度计算所述目标用户和所述候选用户相对于所述元路径的用户相似度包括:
    计算获取的相对于各条原子元路径的相似度的和;及,
    将所述相似度的和直接或者进行正相关运算后作为所述目标用户和所述 候选用户相对于所述元路径的用户相似度。
  20. 根据权利要求17所述的非易失性可读存储介质,其特征在于,所述根据所述候选用户和候选推荐对象间的连接的属性值、所述元路径的属性值约束条件以及所述用户相似度,估计所述候选推荐对象和所述目标用户间的连接的属性值包括:
    获取所述目标用户和候选推荐对象间的连接的属性值的离散取值范围;
    对于所述离散取值范围内的每个取值,分别获取具有与所述取值满足所述属性值约束条件的属性值的所述候选用户和候选推荐对象间的连接,根据获取的连接所对应的候选用户与所述目标用户间的用户相似度计算所述取值对应的属性值强度;
    将所述离散取值范围内的各个取值分别以相应的属性值强度为权重计算加权平均值;及
    根据计算出的加权平均值获得所述候选推荐对象和所述目标用户间的连接的估计的属性值。
  21. 根据权利要求20所述的非易失性可读存储介质,其特征在于,所述根据计算出的加权平均值获得所述候选推荐对象和所述目标用户间的连接的估计的属性值包括:
    将各个元路径下计算出的加权平均值分别乘以相应的元路径的路径权重以计算加权平均值,获得所述候选推荐对象和所述目标用户间的连接的估计的属性值。
  22. 根据权利要求20所述的非易失性可读存储介质,其特征在于,所述根据计算出的加权平均值获得所述候选推荐对象和所述目标用户间的连接的估计的属性值包括:
    将各个元路径下计算出的加权平均值分别乘以与目标用户和相应的元路径对应的路径权重以计算加权平均值,获得所述候选推荐对象和所述目标用户间的连接的估计的属性值。
  23. 根据权利要求22所述的非易失性可读存储介质,其特征在于,还包 括:
    获取所述候选推荐对象和所述目标用户间的连接的真实的属性值;
    将与目标用户和元路径对应的路径权重初始化;及
    根据所述用户相似度,朝趋近于与所述候选用户和所述元路径对应的路径权重的平均值的方向调整初始化的路径权重,使得真实的属性值和估计的属性值的差距满足最小化条件。
  24. 根据权利要求17所述的非易失性可读存储介质,其特征在于,所述候选推荐对象为网络资源;所述属性值为评分值。
PCT/CN2016/084284 2015-09-08 2016-06-01 推送推荐信息的方法、服务器及存储介质 WO2017041541A1 (zh)

Priority Applications (1)

Application Number Priority Date Filing Date Title
US15/715,840 US10609433B2 (en) 2015-09-08 2017-09-26 Recommendation information pushing method, server, and storage medium

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201510567428.9 2015-09-08
CN201510567428.9A CN106503022B (zh) 2015-09-08 2015-09-08 推送推荐信息的方法和装置

Related Child Applications (1)

Application Number Title Priority Date Filing Date
US15/715,840 Continuation US10609433B2 (en) 2015-09-08 2017-09-26 Recommendation information pushing method, server, and storage medium

Publications (1)

Publication Number Publication Date
WO2017041541A1 true WO2017041541A1 (zh) 2017-03-16

Family

ID=58239123

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2016/084284 WO2017041541A1 (zh) 2015-09-08 2016-06-01 推送推荐信息的方法、服务器及存储介质

Country Status (3)

Country Link
US (1) US10609433B2 (zh)
CN (1) CN106503022B (zh)
WO (1) WO2017041541A1 (zh)

Cited By (16)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107358447A (zh) * 2017-06-29 2017-11-17 安徽大学 一种以服务质量为中心的个性化服务推荐方法和***
CN108021657A (zh) * 2017-12-01 2018-05-11 四川大学 一种基于文献标题语义信息的相似作者搜索方法
CN110264306A (zh) * 2019-05-21 2019-09-20 平安银行股份有限公司 基于大数据的产品推荐方法、装置、服务器及介质
CN110377818A (zh) * 2019-06-14 2019-10-25 平安科技(深圳)有限公司 用户信息推送方法、装置、存储介质和计算机设备
CN110427574A (zh) * 2019-08-02 2019-11-08 江苏满运软件科技有限公司 路线相似度确定方法、装置、设备和介质
CN110443493A (zh) * 2019-08-02 2019-11-12 江苏满运软件科技有限公司 路线相似度确定方法、装置、设备和介质
CN110633422A (zh) * 2019-09-16 2019-12-31 安徽大学 一种基于异构信息网络的微博好友推荐方法
CN110969449A (zh) * 2018-09-28 2020-04-07 北京嘀嘀无限科技发展有限公司 车辆尾号的确定方法和装置
CN111708950A (zh) * 2020-06-22 2020-09-25 腾讯科技(深圳)有限公司 内容推荐方法、装置及电子设备
CN111815396A (zh) * 2020-06-04 2020-10-23 清华大学 基于元图的产品筛选方法、***、设备及存储介质
CN111861605A (zh) * 2019-04-28 2020-10-30 阿里巴巴集团控股有限公司 业务对象推荐方法
CN112148822A (zh) * 2020-08-28 2020-12-29 中国地质大学(武汉) 一种细粒度属性加权方法及***
CN112750030A (zh) * 2021-01-11 2021-05-04 深圳前海微众银行股份有限公司 风险模式识别方法、装置、设备及计算机可读存储介质
CN113837847A (zh) * 2021-10-29 2021-12-24 重庆大学 基于异质多元关系融合的知识密集型服务推荐方法
CN115136163A (zh) * 2022-05-25 2022-09-30 广东逸动科技有限公司 管理方法、管理装置、电子设备及计算机可读存储介质
CN116955736A (zh) * 2023-09-15 2023-10-27 北京南天智联信息科技股份有限公司 数据标准中数据约束条件推荐方法及***

Families Citing this family (31)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107103057B (zh) 2017-04-13 2018-09-18 腾讯科技(深圳)有限公司 一种资源推送方法及装置
CN107239993B (zh) * 2017-05-24 2020-11-24 海南大学 一种基于拓展标签的矩阵分解推荐方法及***
CN109428928B (zh) * 2017-08-31 2021-01-05 腾讯科技(深圳)有限公司 信息推送对象的选择方法、装置及设备
CN109429178B (zh) * 2017-09-04 2021-06-25 腾讯科技(北京)有限公司 信息推送方法、装置、存储介质和电子装置
CN107944629B (zh) * 2017-11-30 2020-08-07 北京邮电大学 一种基于异质信息网络表示的推荐方法及装置
CN108805628B (zh) * 2018-06-19 2021-03-16 东北师范大学 一种电子商务推荐方法
CN108985830B (zh) * 2018-07-05 2021-03-02 北京邮电大学 基于异质信息网络的推荐评分方法、装置
CN109308315B (zh) * 2018-10-19 2022-09-16 南京理工大学 一种基于专家领域相似度与关联关系的协同推荐方法
CN109522483B (zh) * 2018-11-14 2022-04-12 北京百度网讯科技有限公司 用于推送信息的方法和装置
CN111292107A (zh) * 2018-12-06 2020-06-16 北京嘀嘀无限科技发展有限公司 用户数据处理方法、装置、电子设备及存储介质
CN109685630B (zh) * 2019-01-09 2020-10-27 苏州大学 电子商务团购推荐方法及***
CN109800232B (zh) * 2019-01-21 2021-03-19 北京邮电大学 一种异质信息网络嵌入方法、装置、电子设备及存储介质
CN110009455B (zh) * 2019-04-02 2022-02-15 长安大学 一种基于网络表示学习的网约共享出行人员匹配方法
CN110083766B (zh) * 2019-04-26 2021-05-14 北京邮电大学 一种基于元路径引导嵌入的查询推荐方法及装置
CN110598124B (zh) * 2019-08-14 2023-05-26 平安科技(深圳)有限公司 数值属性挖掘方法、装置、计算机设备和存储介质
CN110609889B (zh) * 2019-08-30 2021-11-23 中国科学院计算技术研究所 基于学术网络确定对象重要性排名及选择评审专家的方法和***
CN110689410B (zh) * 2019-09-29 2021-09-03 京东数字科技控股有限公司 数据处理方法、装置、设备及存储介质
CN110930223A (zh) * 2019-11-21 2020-03-27 苏州朗动网络科技有限公司 基于场感知因子分解机的推荐召回方法、设备和存储介质
CN111507783B (zh) * 2020-04-13 2023-05-02 卓尔智联(武汉)研究院有限公司 电子票据推送方法、装置、区块链节点设备和存储介质
CN111475637B (zh) * 2020-06-24 2020-10-09 支付宝(杭州)信息技术有限公司 用于推送知识点的数据处理、训练方法及装置
CN111832724B (zh) * 2020-07-14 2023-03-28 西北工业大学 一种基于深度神经网络的异构网络推荐方法
CN112214683B (zh) * 2020-09-09 2024-05-14 华南师范大学 基于异构信息网络的混合推荐模型处理方法、***和介质
CN112131480B (zh) * 2020-09-30 2022-06-10 中国海洋大学 基于多层异质属性网络表征学习的个性化商品推荐方法及***
CN112232492B (zh) * 2020-10-30 2022-04-12 北京邮电大学 一种基于解耦的异质网络嵌入方法、装置及电子设备
CN112380494B (zh) * 2020-11-17 2023-09-01 ***股份有限公司 一种确定对象特征的方法及装置
CN112507213B (zh) * 2020-11-26 2022-09-30 杭州讯酷科技有限公司 一种基于行为大数据分析的推荐优化的***方案的方法
CN112650944A (zh) * 2020-12-29 2021-04-13 哈尔滨理工大学 一种基于高阶连接图的协同过滤推荐算法
CN113239266B (zh) * 2021-04-07 2023-03-14 中国人民解放军战略支援部队信息工程大学 基于局部矩阵分解的个性化推荐方法及***
CN113192626B (zh) * 2021-04-13 2022-09-13 山东大学 基于孪生神经网络的用药方案推荐***及方法
CN113742573B (zh) * 2021-08-03 2023-11-14 深圳Tcl新技术有限公司 一种多媒体资源搜索方法、装置、计算机设备和存储介质
CN114757745A (zh) * 2022-04-25 2022-07-15 北京航空航天大学 一种基于联合矩阵分解的电商场景挖掘方法与***

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103136289A (zh) * 2011-12-05 2013-06-05 腾讯科技(深圳)有限公司 资源推荐方法及***
CN103955535A (zh) * 2014-05-14 2014-07-30 南京大学镇江高新技术研究院 一种基于元路径的个性化推荐方法及***
CN104077351A (zh) * 2014-05-26 2014-10-01 东北师范大学 基于异构信息网络的内容提供方法及***
CN104239496A (zh) * 2014-09-10 2014-12-24 西安电子科技大学 一种结合模糊权重相似性度量和聚类协同过滤的方法
CN104794636A (zh) * 2015-04-17 2015-07-22 南京大学 基于用户显示评分的机型推荐方法
CN104850632A (zh) * 2015-05-22 2015-08-19 东北师范大学 一种通用的基于异构信息网络的相似度计算方法与***

Family Cites Families (12)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US5758257A (en) * 1994-11-29 1998-05-26 Herz; Frederick System and method for scheduling broadcast of and access to video programs and other data using customer profiles
US8341158B2 (en) * 2005-11-21 2012-12-25 Sony Corporation User's preference prediction from collective rating data
US8738695B2 (en) * 2007-05-15 2014-05-27 International Business Machines Corporation Joint analysis of social and content networks
US9185469B2 (en) * 2010-09-30 2015-11-10 Kodak Alaris Inc. Summarizing image collection using a social network
US20120324491A1 (en) * 2011-06-17 2012-12-20 Microsoft Corporation Video highlight identification based on environmental sensing
US20160350834A1 (en) * 2015-06-01 2016-12-01 Nara Logics, Inc. Systems and methods for constructing and applying synaptic networks
US8984082B2 (en) * 2012-08-29 2015-03-17 Wetpaint.Com, Inc. Personalization based upon social value in online media
US9686577B2 (en) * 2013-02-22 2017-06-20 Facebook Time-sensitive content update
CN104035934B (zh) * 2013-03-06 2019-01-15 腾讯科技(深圳)有限公司 一种多媒体信息推荐的方法及装置
CN103336793B (zh) * 2013-06-09 2015-08-12 中国科学院计算技术研究所 一种个性化论文推荐方法及其***
US9552055B2 (en) * 2013-07-15 2017-01-24 Facebook, Inc. Large scale page recommendations on online social networks
US20150149205A1 (en) * 2013-11-27 2015-05-28 University-Industry Cooperation Group Of Kyung Hee University Integrated clinical decision supporting system and method

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103136289A (zh) * 2011-12-05 2013-06-05 腾讯科技(深圳)有限公司 资源推荐方法及***
CN103955535A (zh) * 2014-05-14 2014-07-30 南京大学镇江高新技术研究院 一种基于元路径的个性化推荐方法及***
CN104077351A (zh) * 2014-05-26 2014-10-01 东北师范大学 基于异构信息网络的内容提供方法及***
CN104239496A (zh) * 2014-09-10 2014-12-24 西安电子科技大学 一种结合模糊权重相似性度量和聚类协同过滤的方法
CN104794636A (zh) * 2015-04-17 2015-07-22 南京大学 基于用户显示评分的机型推荐方法
CN104850632A (zh) * 2015-05-22 2015-08-19 东北师范大学 一种通用的基于异构信息网络的相似度计算方法与***

Cited By (28)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107358447A (zh) * 2017-06-29 2017-11-17 安徽大学 一种以服务质量为中心的个性化服务推荐方法和***
CN107358447B (zh) * 2017-06-29 2021-01-29 安徽大学 一种以服务质量为中心的个性化服务推荐方法和***
CN108021657A (zh) * 2017-12-01 2018-05-11 四川大学 一种基于文献标题语义信息的相似作者搜索方法
CN110969449A (zh) * 2018-09-28 2020-04-07 北京嘀嘀无限科技发展有限公司 车辆尾号的确定方法和装置
CN110969449B (zh) * 2018-09-28 2023-12-05 北京嘀嘀无限科技发展有限公司 车辆尾号的确定方法和装置
CN111861605A (zh) * 2019-04-28 2020-10-30 阿里巴巴集团控股有限公司 业务对象推荐方法
CN110264306A (zh) * 2019-05-21 2019-09-20 平安银行股份有限公司 基于大数据的产品推荐方法、装置、服务器及介质
CN110264306B (zh) * 2019-05-21 2023-06-30 平安银行股份有限公司 基于大数据的产品推荐方法、装置、服务器及介质
CN110377818A (zh) * 2019-06-14 2019-10-25 平安科技(深圳)有限公司 用户信息推送方法、装置、存储介质和计算机设备
CN110427574A (zh) * 2019-08-02 2019-11-08 江苏满运软件科技有限公司 路线相似度确定方法、装置、设备和介质
CN110443493A (zh) * 2019-08-02 2019-11-12 江苏满运软件科技有限公司 路线相似度确定方法、装置、设备和介质
CN110443493B (zh) * 2019-08-02 2022-09-23 贵阳货车帮科技有限公司 路线相似度确定方法、装置、设备和介质
CN110633422A (zh) * 2019-09-16 2019-12-31 安徽大学 一种基于异构信息网络的微博好友推荐方法
CN110633422B (zh) * 2019-09-16 2022-04-05 安徽大学 一种基于异构信息网络的微博好友推荐方法
CN111815396A (zh) * 2020-06-04 2020-10-23 清华大学 基于元图的产品筛选方法、***、设备及存储介质
CN111815396B (zh) * 2020-06-04 2024-05-31 清华大学 基于元图的产品筛选方法、***、设备及存储介质
CN111708950B (zh) * 2020-06-22 2023-08-29 腾讯科技(深圳)有限公司 内容推荐方法、装置及电子设备
CN111708950A (zh) * 2020-06-22 2020-09-25 腾讯科技(深圳)有限公司 内容推荐方法、装置及电子设备
CN112148822A (zh) * 2020-08-28 2020-12-29 中国地质大学(武汉) 一种细粒度属性加权方法及***
CN112148822B (zh) * 2020-08-28 2024-04-19 中国地质大学(武汉) 一种细粒度属性加权方法及***
CN112750030A (zh) * 2021-01-11 2021-05-04 深圳前海微众银行股份有限公司 风险模式识别方法、装置、设备及计算机可读存储介质
CN112750030B (zh) * 2021-01-11 2024-04-26 深圳前海微众银行股份有限公司 风险模式识别方法、装置、设备及计算机可读存储介质
CN113837847A (zh) * 2021-10-29 2021-12-24 重庆大学 基于异质多元关系融合的知识密集型服务推荐方法
CN113837847B (zh) * 2021-10-29 2023-08-29 重庆大学 基于异质多元关系融合的知识密集型服务推荐方法
CN115136163A (zh) * 2022-05-25 2022-09-30 广东逸动科技有限公司 管理方法、管理装置、电子设备及计算机可读存储介质
CN115136163B (zh) * 2022-05-25 2023-08-18 广东逸动科技有限公司 管理方法、管理装置、电子设备及计算机可读存储介质
CN116955736A (zh) * 2023-09-15 2023-10-27 北京南天智联信息科技股份有限公司 数据标准中数据约束条件推荐方法及***
CN116955736B (zh) * 2023-09-15 2023-12-01 北京南天智联信息科技股份有限公司 数据标准中数据约束条件推荐方法及***

Also Published As

Publication number Publication date
CN106503022A (zh) 2017-03-15
US20180020250A1 (en) 2018-01-18
CN106503022B (zh) 2020-12-01
US10609433B2 (en) 2020-03-31

Similar Documents

Publication Publication Date Title
WO2017041541A1 (zh) 推送推荐信息的方法、服务器及存储介质
CN108804633B (zh) 基于行为语义知识网络的内容推荐方法
Zhang et al. Enabling kernel-based attribute-aware matrix factorization for rating prediction
EP4181026A1 (en) Recommendation model training method and apparatus, recommendation method and apparatus, and computer-readable medium
WO2018090545A1 (zh) 融合时间因素的协同过滤方法、装置、服务器和存储介质
WO2020233432A1 (zh) 一种信息推荐方法及装置
WO2018232331A1 (en) Systems and methods for optimizing and simulating webpage ranking and traffic
Bojchevski et al. Is pagerank all you need for scalable graph neural networks
JP6417688B2 (ja) キュレーションをランク付けする方法及びシステム
Duan et al. A hybrid intelligent service recommendation by latent semantics and explicit ratings
WO2022111095A1 (zh) 一种产品推荐方法、装置、计算机存储介质及***
Yin et al. An efficient recommendation algorithm based on heterogeneous information network
Liang et al. Measure prediction capability of data for collaborative filtering
Qian et al. Three-way decision collaborative recommendation algorithm based on user reputation
Li et al. Session Recommendation Model Based on Context‐Aware and Gated Graph Neural Networks
Guan et al. Enhanced SVD for collaborative filtering
CN113657116B (zh) 基于视觉语义关系的社交媒体流行度预测方法及装置
Liu et al. Speak the same language with your friends: augmenting tag recommenders with social relations
Liu et al. Adaptive local low-rank matrix approximation for recommendation
WO2014007943A2 (en) Method and apparatus for obfuscating user demographics
CN116340645A (zh) 基于图增强协同过滤的内容推荐
Ting et al. Transfer-learning based model for reciprocal recommendation
JP5938006B2 (ja) アイテム推薦装置、アイテム推薦方法、およびアイテム推薦プログラム
Felıcio et al. Social prefrec framework: leveraging recommender systems based on social information
Sharma et al. Personalized Fashion Recommendation Using Nearest Neighbor PageRank Algorithm

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: 16843479

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 16843479

Country of ref document: EP

Kind code of ref document: A1