CN106649509B

CN106649509B - User feature extraction method and device

Info

Publication number: CN106649509B
Application number: CN201610891915.5A
Authority: CN
Inventors: 李会珠; 卫磊
Original assignee: Tencent Technology Shenzhen Co Ltd
Current assignee: Tencent Technology Shenzhen Co Ltd
Priority date: 2016-10-12
Filing date: 2016-10-12
Publication date: 2020-04-07
Anticipated expiration: 2036-10-12
Also published as: CN106649509A

Abstract

The embodiment of the invention discloses a user feature extraction method and a device, wherein the method comprises the following steps: acquiring user operation behavior data; analyzing the operation behavior data to generate user characteristics; determining an effective time window of the user characteristics according to the operation time window of the user operation behavior data; wherein the user characteristic is used for controlling the recommendation probability of the first type of information in the effective time window.

Description

User feature extraction method and device

Technical Field

The invention relates to the technical field of information, in particular to a user feature extraction method and device.

Background

With the development of information technology, the reading application service is more and more intelligent. For example, through the extraction of the user characteristics, only interested information is recommended to the user and/or uninteresting information is prevented from being recommended to the user, so that the individual requirements of the user are met, the user is prevented from searching the interested information in massive information, the software and hardware resources of the electronic equipment are better utilized, and the reading of the user is optimized. To implement the personalized recommendation of the user, the accuracy of the user feature extraction is very important. In the prior art, a plurality of methods for extracting user features are provided, but when information recommendation is performed by using the user features extracted by the methods, the expected accuracy cannot be met. Therefore, how to improve the accuracy of user feature extraction is a further technical problem to be solved urgently in the prior art.

Disclosure of Invention

In view of this, embodiments of the present invention are intended to provide a method and an apparatus for extracting user features, which are used to solve, at least in part, the problem of inaccurate user feature extraction.

In order to achieve the purpose, the technical scheme of the invention is realized as follows:

a first aspect of an embodiment of the present invention provides a user feature extraction method, including:

acquiring user operation behavior data;

analyzing the operation behavior data to generate user characteristics;

determining an effective time window of the user characteristics according to the operation time window of the user operation behavior data;

wherein the user characteristic is used for controlling the recommendation probability of the first type of information in the effective time window.

Based on the above scheme, the analyzing the operation behavior data to generate the user characteristics includes:

generating short-term user features based on the operational behavior data within the first operational time window and generating long-term user features based on the operational behavior data within a second operational time window; the duration of the second operation time window is greater than the duration of the first operation time window;

the determining an effective time window of the user characteristics according to the operation time window of the user operation behavior data includes:

generating a first effective time window of the short-term user characteristics and a second effective time window of the long-term user characteristics according to the duration of the operation time window; wherein the duration of the second validity time window is greater than the duration of the first validity time window.

Based on the above scheme, the generating long-term user characteristics based on the operation behavior data in the second operation time window includes:

analyzing the operation behavior data to obtain the occurrence frequency of the potential long-term features in the potential long-term feature set;

determining whether a reverse feature corresponding to a first potential long-term feature included in the potential long-term feature set is a formal user feature in the formal user feature set;

determining that the first potential long-term feature is the long-term user feature if the first potential long-term feature is not the official user feature.

Based on the above scheme, the generating long-term user characteristics based on the operation behavior data in the second operation time window further includes:

if the first potential long-term feature is a first type of formal user feature in the formal user feature set, deleting the first potential long-term feature from the potential long-term feature set; the first type of formal user characteristics are formal user characteristics with the ranking value in a preset range; the scoring value is used for controlling the recommendation probability of the information corresponding to the formal user characteristics.

if the first potential long-term feature is a second type of formal user feature in the formal user feature set, adjusting the score value of the formal user feature corresponding to the first potential long-term feature; the second type of formal user characteristics are characteristics of which the score value is outside the predetermined range.

counting the occurrence frequency of the short-term user characteristics in the second operation time window;

when the frequency of occurrence is greater than a predetermined threshold, converting the short-term user characteristics to the long-term user characteristics.

Based on the above scheme, the generating short-term user features based on the operation behavior data in the first operation time window, and generating long-term user features based on the operation behavior data in the second operation time window further includes:

analyzing negative operation behavior data in the first operation time window to generate short-term negative characteristics; analyzing negative operation behavior data in the second operation time window to generate long-term negative characteristics;

wherein the negative-going operational behavior characterized by the negative-going operational behavior data is an operation that reduces display of information.

Based on the above scheme, the negative operation behavior includes: a delete operation and/or a first type close operation; the first type of closing operation comprises closing operation of information links and/or closing operation of which detailed page display time is less than a first preset time length; the information link is: and entering a link of a detailed page of the detailed information display corresponding to the information link.

A second aspect of the embodiments of the present invention provides a user feature extraction device, including:

the acquisition unit is used for acquiring user operation behavior data;

the analysis unit is used for analyzing the operation behavior data to generate user characteristics;

the determining unit is used for determining an effective time window of the user characteristics according to the operation time window in which the user operation behavior data is positioned;

Based on the above scheme, the parsing unit is configured to generate a short-term user feature based on the operation behavior data in the first operation time window, and generate a long-term user feature based on the operation behavior data in the second operation time window; the duration of the second operation time window is greater than the duration of the first operation time window;

the determining unit is specifically configured to generate a first effective time window of the short-term user characteristic and a second effective time window of the long-term user characteristic according to the duration of the operation time window; wherein the duration of the second validity time window is greater than the duration of the first validity time window.

Based on the above scheme, the analyzing unit is specifically configured to analyze the operation behavior data to obtain the occurrence frequency of the potential long-term features in the potential long-term feature set; determining whether a reverse feature corresponding to a first potential long-term feature included in the potential long-term feature set is a formal user feature in the formal user feature set; and if the reverse characteristic of the first potential long-term characteristic is not the formal user characteristic, determining that the first potential characteristic is the long-term user characteristic according to the occurrence frequency.

Based on the foregoing scheme, the parsing unit is further specifically configured to delete the first potential long-term feature from the potential long-term feature set if the reverse feature of the first potential long-term feature is a first type of official user feature in the official user feature set; the first type of formal user characteristics are formal user characteristics with the ranking value in a preset range; the scoring value is used for controlling the recommendation probability of the information corresponding to the formal user characteristics.

Based on the above scheme, the parsing unit is further specifically configured to adjust a score value of a reverse feature of the first potential long-term feature if the first potential long-term feature is a second type of official user feature in the official user feature set; the second type of formal user characteristics are characteristics of which the score value is outside the predetermined range.

Based on the above scheme, the parsing unit is specifically configured to count occurrence frequencies of the short-term user features in the second operation time window; when the frequency of occurrence is greater than a predetermined threshold, converting the short-term user characteristics to the long-term user characteristics.

Based on the above scheme, the analyzing unit is further configured to analyze negative operation behavior data in the first operation time window to generate a short-term negative feature; analyzing negative operation behavior data in the second operation time window to generate long-term negative characteristics; wherein the negative-going operational behavior characterized by the negative-going operational behavior data is an operation that reduces display of information.

According to the user feature extraction method and device provided by the embodiment of the invention, the user behavior data can be obtained, and the effective time window is determined according to the operation time window corresponding to the user behavior data. Clearly, this avoids the phenomenon of being used as a permanent user feature once it is generated, relative to the prior art. Because the user's preferences may change and it is difficult to achieve a hundred percent accuracy in the extraction of the user's features. If a feature is extracted as a permanent feature once, the feature of the user is inaccurate for a long time as time goes on or if the feature extraction is biased, and if the processing such as user information pushing continues according to the inaccurate feature of the user, personalized information pushing of the user can be realized, but user's dislike is incurred, and data which the user is interested in cannot be effectively propagated. In the embodiment, the effective time window is set for the user characteristics according to the operation time window, so that the problems can be well solved, and the accuracy of the user characteristics is improved.

Drawings

Fig. 1 is a schematic flow chart of a first user feature extraction method according to an embodiment of the present invention;

FIG. 2 is a schematic diagram of an information link according to an embodiment of the present invention;

FIG. 3 is a schematic diagram illustrating a detailed page of information according to an embodiment of the present invention;

FIG. 4 is a schematic illustration of a display of another detailed page of information provided by an embodiment of the present invention;

FIG. 5 is a flowchart illustrating a short-term user feature extraction according to an embodiment of the present invention;

fig. 6 is a schematic structural diagram of a user feature extraction apparatus according to an embodiment of the present invention;

fig. 7 is a schematic structural diagram of an information framework according to an embodiment of the present invention;

fig. 8 is a schematic flowchart of another user feature extraction method according to an embodiment of the present invention;

fig. 9 is a schematic structural diagram of another user feature extraction apparatus according to an embodiment of the present invention.

Detailed Description

The technical solution of the present invention is further described in detail with reference to the drawings and the specific embodiments of the specification.

As shown in fig. 1, the present embodiment provides a user feature extraction method, including:

step S110: acquiring user operation behavior data;

step S120: analyzing the operation behavior data to generate user characteristics;

step S130: determining an effective time window of the user characteristics according to the operation time window of the user operation behavior data;

The user feature extraction method described in this embodiment may be applied to electronic devices such as various servers with data processing.

The step S110 of obtaining the user behavior data may include obtaining the user behavior data from a Business and operation Support System (Boss) System, or receiving the user behavior data reported by a client.

The user behavior data may include data resulting from various operations that operate on various information displayed by the client. For example, the user behavior data of various user operation behaviors such as click data, reading behavior data, closing behavior data, deleting behavior data, adding blacklist behavior data, subscribing behavior data, and the like.

Acquiring the user operation behavior data according to a preset time interval; wherein the first predetermined time includes at least two of the predetermined time intervals.

The step S110 may include: the user operation behavior data is acquired according to a preset time interval, wherein the preset time interval can be an equal preset time interval, namely the user operation behavior data is acquired periodically, or can be an unequal preset time interval, namely the user operation behavior data is acquired non-periodically. In short, in this embodiment, the user operation behavior data is acquired at predetermined time intervals, and conversion from the short-term user characteristics to the long-term user characteristics is realized according to the user operation behavior data at a plurality of predetermined time intervals, so as to realize the accuracy of the user characteristics.

The step S120 may include: and analyzing whether preset operation is executed or not, and information such as the execution times or frequency of the preset operation is obtained in the operation time window.

In this embodiment, the method further includes step S130, where step S130 generates a failure time window corresponding to the user characteristic according to the operation time window. Typically, the duration of the validation time window is proportional to the duration of the operational time window. Since the longer the operation time window is, the more the corresponding user operation behavior data is, the more stable preference of the user can be reflected, and the higher the accuracy of the reflection is, the duration of the effective time window and the duration of the operation time window generated in the embodiment are.

In the embodiment, after the user features are extracted, a concept of an effective time window is also introduced, and after the effective time window is effective, the user features are invalid. For example, the corresponding user features are cleared in the database, so that if the user attributes such as the preference of the user change, the failure of the user features can avoid the problem of inaccuracy caused by the consistent effect of the user features, and the accuracy of user feature extraction is improved.

For example, the step S120 may include:

the step S130 may include:

In this embodiment, the user characteristics include at least two types, a short-term user characteristic and a long-term user characteristic. The mood of the user may also change, for example, some information is interested in a certain time, but after the time period, the interest of the user is changed back. However, the user operation behavior data corresponding to the user feature is not considered as long-term user feature extraction in the second operation time window. In order to take account of the long-term preference and the short-term preference of the user, at least two user characteristics, namely a short-term user characteristic and a long-term user characteristic, are generated in the embodiment, and the effective time of the short-term user characteristic is shorter. Once the user characteristics fail, the information recommendation probability determined according to the user characteristics needs to be confirmed again.

The short-term user characteristics may be used to characterize the user's current preferences. The effective time window of the short-term user behavior characteristic action is the first effective time window. The first life time window may be 24 hours, a predetermined number of days, a week, a month, or the like.

The long-term user characteristic may be that a duration of a second validity time window is greater than the first validity time window.

In short, the embodiment provides the user feature extraction method which can give consideration to both short-term preference and long-term preference of the user, and the accuracy of the user features is improved again.

In a specific implementation process, the step S120 may include: extracting short-term user features by using a first analysis mode, and extracting long-term user features by using a second analysis mode; the first analysis manner is different from the second analysis manner. In this embodiment, the first analysis manner and the second analysis manner may be set according to characteristics of the short-term user characteristics and the long-term user characteristics. For example, due to the effective duration of the short-term user features, in order to reduce the complexity of data processing and reduce the data processing amount, the complexity of the first analysis method may be lower than the complexity of the second analysis method, and the complexity may represent the number of operation steps of data analysis, the introduced data processing objects of different dimensions or types, and/or the calculation amount of a single step. Therefore, the individual requirements of the user characteristics with different effective time lengths can be well met, for example, the calculation amount and the accuracy requirement of the user characteristics are well balanced.

The following respectively introduces the short-term user feature and the long-term user feature extraction methods:

generation of short-term user features:

the step S120 may include:

analyzing the user behavior data in the first operation time window, and determining the occurrence frequency of each operation behavior;

and determining the short-term user characteristics according to the frequency of occurrence.

For example, the user behavior data is clustered by using a clustering method. If the frequency of occurrence is greater than a particular threshold, the short-term user profile may be generated, and one or more of the operational behaviors having the highest frequency of occurrence may be selected to generate the short-term user profile. There are various ways to extract the short-term user features, and the short-term user features are not limited to any of the above.

Extracting long-term user features:

the step S120 may include:

analyzing the operation behavior data, and determining the occurrence frequency of operation behaviors corresponding to the potential long-term features in the potential long-term feature set;

and if the reverse characteristic of the first potential long-term characteristic is not the formal user characteristic, determining that the first potential characteristic is the long-term user characteristic according to the occurrence frequency.

The formal user feature set here is a user feature determined at a historical time. The inverse characteristic of the first potential long-term characteristic is a characteristic that characterizes the user preference as being opposite to the first potential long-term characteristic. For example, the potential long-term characteristic is that the user does not like class a information, the reverse characteristic of the potential long-term characteristic is that the user likes class a information, and if the characteristic that the user likes class a information does not exist in the current formal user characteristic set, it is obvious that the extracted long-term user characteristic represents that the user does not like class a information in the second operation time window. The features that the user likes the type a information and the features that the user dislikes the type a information are mutually reverse features.

And if one potential long-term characteristic is not in the formal user characteristic, determining whether the formal long-term user characteristic is over according to the occurrence frequency. For example, the determination of the formal long-term user characteristics may include: and analyzing the operation behavior data to obtain the occurrence frequency of a certain operation behavior, and considering the operation behavior as a potential long-term characteristic when the occurrence frequency meets a preset condition. For example, if the frequency of occurrence is greater than a threshold or the frequency of occurrence is ranked first, both may be considered to satisfy the preset condition, then it may be confirmed that the long-term user characteristic is formal.

Further, the step S120 further includes:

The formal user features may be divided into at least two categories. The first type of formal user characteristics are characteristics with the scoring value within a preset range. In the embodiment, the level of the score value is positively correlated with the occurrence frequency of the operation behavior for extracting and generating the corresponding user feature. For example, if the score value is higher, the recommendation probability is higher, and if the frequency of occurrence of a predetermined operation behavior is higher, the score value of the corresponding user feature is higher. Of course, in a specific implementation, the lower the score value is, the lower the recommendation probability is, and if the frequency of occurrence of a predetermined operation behavior is higher, the score value of the corresponding user feature is lower. In this embodiment, the first type of user features are user features whose corresponding recommendation probability is greater than a threshold or whose recommendation probability is ranked first, and the recommendation probability at this time is ranked from high to low.

If the reverse feature of a potential long-term feature is the first type of formal feature, the current operation behavior for generating the potential long-term feature may be a user's misoperation, and therefore, in order to eliminate the extraction of the user feature of the misoperation, the potential long-term feature is deleted from the corresponding set in the present embodiment, so as to improve the correctness of the user feature.

In other embodiments, step S120 further includes:

if the reverse feature of the first potential long-term feature is a second type of official user feature in the official user feature set, adjusting the score value of the reverse feature of the first potential long-term feature; the second type of formal user characteristics are characteristics of which the score value is outside the predetermined range.

The second type of formal user features are user features other than the first type of formal user features. In order to adjust the corresponding recommendation probability in this embodiment, the score value of the corresponding reverse feature is adjusted in this embodiment, so as to adjust the recommendation probability.

For example, in this embodiment, the second type of formal user features are features other than the first type of formal user features, that is, the degree that the user likes a certain type of information or performs a certain operation is relatively low. Therefore, in this embodiment, the first formal feature corresponding to the first potential long-term feature is the second type of forward feature, and the score value is adjusted to reduce the recommendation probability of the information corresponding to the first forward feature. For example, if the score value is higher, it indicates that the user likes to a higher degree; the score value of the second type of forward feature is relatively lower than the score value of the first type of forward feature, and the score value is adjusted to be lower in this embodiment. In this embodiment, the score value is a recommendation probability for recommending a certain type of information to a user. Therefore, in the embodiment, the score value is adjusted to reduce the recommendation probability; in addition, in the process of determining the long-term user characteristics, the adjustment in the scoring of the positive user characteristics is realized through the adjustment of the scoring value, so that the re-correction of the recommendation probability is realized.

The step S120 further includes: counting the occurrence frequency of the short-term user characteristics in the second operation time window; when the frequency of occurrence is greater than a predetermined threshold, converting the short-term user characteristics to the long-term user characteristics. For example, if a short-term user feature frequently appears in the second operation time window, the short-term user feature may be a stable feature of the user and may be converted into the long-term user feature. In the embodiment, the long-term user features are determined by using the processing result of the short-term user features, so that the extraction complexity of the long-term user features can be reduced, and the calculation amount is saved.

In this embodiment, the occurrence frequency of the short-term user feature in the second operation time is continuously counted, and finally, whether to convert the short-term user feature into the long-term user feature is determined according to the occurrence frequency. For example, the short-term user characteristics are determined with 24 hours as a statistical period. For example, the short-term user characteristic a characterizes that the user does not like class a information. The first predetermined time may be 3 days, if the frequency of the short-term user characteristics a is lower than the predetermined threshold within 3 days, the conversion of the short-term user characteristics a is not performed, and if the frequency of the short-term user characteristics a is greater than the predetermined threshold, the short-term user characteristics a are converted into the long-term user characteristics a. And prohibiting the recommendation of the type A information within a week according to the short-term user characteristic A, and prohibiting the recommendation of the type A information within a month according to the long-term user characteristic A. Here, one week may be the first effective time window, and one month may be the second effective time window. After the effective time window of the short-term user characteristics and the long-term user characteristics is invalid, the recommendation of the information corresponding to the user characteristics can be tried again to extract the user characteristics again, so that the phenomenon that the user characteristics are not applicable any more due to the change of the user preferences is avoided.

The user features in this embodiment may include positive user features and negative user features. The forward user characteristics are characteristics which represent that a user likes a certain type of information; the negative user characteristics represent characteristics that a user dislikes or dislikes a certain kind of information.

The step S120 may include:

analyzing negative operation behavior data in the first operation time window to generate short-term negative characteristics; analyzing negative operation behavior data in the second operation time window to generate long-term negative characteristics; wherein the negative-going operational behavior characterized by the negative-going operational behavior data is an operation that reduces display of information.

The long-term negative-going feature and the short-term negative-going feature are both one of the aforementioned negative-going user features, differing by the validation time window.

The negative operation may include: a delete operation and/or a first type close operation; the first type of closing operation comprises closing operation of information links and/or closing operation of which detailed page display time is less than a first preset time length; the information link is: and entering a link of a detailed page of the detailed information display corresponding to the information link.

For example, in an information display platform of a social application, certain information is displayed; the user performs a masking operation of the information, a closing operation of the information display, deletion of the information, or the like to reduce negative operations of the information display. In short, the reverse operation is that the user reduces the display of certain information through active operation.

As shown in fig. 2, an information link to a detail page is displayed on the first information display page, and the user directly deletes the information link or closes the information link. The information link can be an information link formed by a title or an abstract of a certain information and entering a detailed page. The information links can be divided into text links and picture links, and the specific implementation process also includes graphic links integrating text and pictures. In fig. 2, the text links include a text link a, a text link b, and a text link c. The picture link comprises a picture link a, a picture link b and a picture link c. Fig. 3 can be regarded as a detailed page, and it is obvious that the detailed page shown in fig. 3 is a display page of the main body or main content of a certain message and information in information links such as a title and/or an abstract.

For another example, although the user opens an information link and then enters a certain detailed page, and a certain piece of information is displayed in the detailed page, the user closes the detailed page soon after entering the page, and it is obvious that the user is not interested in the displayed content in the detailed page. This quick shut down operation is also one of the reverse operations previously described. For example, in the process of information pushing, the user shields the public number a, and then the probability of pushing other public numbers with the same information as the public number a to the user can be reduced. The operation of masking the public number a here is one of the aforementioned reverse operations.

As shown in FIG. 4, after entering the detail page, the display of the close control and timing information is displayed on the detail page. In fig. 4 the close control and timing information are both shown in the upper right hand corner of the detail page. If the detail page is directly forked off or exited within 1 second, the main page as shown in FIG. 2 is entered. In any case, the operation of closing the information page is performed for the purpose of forking off the detail page or exiting the detail page.

In some embodiments, the information analogy feed application has pushed class B information to the user's client N times a week. N1 class B messages are pushed the first day in the push week, and the user closes m1 times directly without entering a detailed page, where m1 is smaller than the positive integer of n 1. By the closing operation of step S110 and step S120 according to the class B message, a short-term negative-going feature that the user dislikes the class B message is generated. The push frequency of the B-type information is reduced from the second day to the fifth day, but push still exists. By turning off the class B messages within the second to fifth days, and with the frequency of turning off being higher or higher, the short-term negative features that the user dislikes the class B messages can be converted into the long-term negative features later in the week. The information push application in the embodiment may include a reading application or a social application with information push. The reading application can comprise various news pushing applications, comment pushing applications and other applications, and the social application can be various applications with social functions such as WeChat or QQ.

In summary, in this embodiment, the short-term negative-direction feature and the long-term negative-direction feature are generated when the user operation behavior data is performed. And the short-term negative-going features and the long-term negative-going features, the time length of the failure time window is inconsistent, so that the problem of insufficient accuracy of the reverse features caused by the fact that the reverse features are used as permanent features once being determined is solved. In addition, the reverse features are divided into short-term negative features and long-term negative features, on one hand, the reverse features are used for processing subsequent information pushing in advance currently through feature extraction, and on the other hand, in the application process of the reverse features, the long-term negative features are determined based on the short-term negative features, so that the determination operation, the determination flow and the determination basis of the long-term negative features are equivalently prolonged, the dimensionality of the reference parameters of the long-term negative features is ensured, and the determination accuracy of the long-term negative features is improved.

In some embodiments, as shown in fig. 5, the step S120 may include:

step S121: analyzing the user operation behavior data and determining an information object corresponding to the user operation behavior;

step S122: extracting information characteristics of the information object to generate an information label;

step S123: analyzing the user operation behavior data to obtain the behavior type and the behavior parameters of the user operation behavior;

step S124: and generating a user characteristic representing whether the information corresponding to the information tag is interesting or not by the user based on the behavior type and the behavior parameters.

In this embodiment, the user operation behavior data may include an operation object, a behavior type, and other behavior parameters. The operation object herein may include the information object. The behavior types may be divided into reverse operation and forward operation. The reverse operation may be an operation to reduce information display; the forward operation is the operation that the user actively reads certain type of information or adds certain type of information. The reading here is for example staying on the detail page for a time exceeding a time threshold. The operation of adding a certain type of information may include an operation of forwarding the type of information or the information to a friend, or an operation of collecting the type of information. The behavior parameters may include a time length for reading a certain message, and a specific operation on the message, for example, an operation for selecting a certain character or copying a certain message, and the like.

As shown in fig. 6, the present embodiment provides a user feature extraction apparatus, including:

an obtaining unit 110, configured to obtain user operation behavior data;

the analyzing unit 120 is configured to analyze the operation behavior data to generate a user characteristic;

a determining unit 130, configured to determine an effective time window of the user characteristic according to an operation time window in which the user operation behavior data is located;

The user feature extraction device provided in this embodiment can be applied to various electronic devices or servers capable of extracting user features.

The acquisition unit 110, the parsing unit 120, and the determination unit 130 may correspond to a processor or a processing circuit. The processor may comprise a central processing unit CPU, a microprocessor MCU, a digital signal processor DSP, an application processor AP, a programmable array PLC, or the like. The processing circuitry may comprise an application specific integrated circuit, ASIC. The processor or processing circuitry may implement the operations of the functional units described above through execution of executable code.

In this embodiment, the user feature is extracted and an effective time window is determined, so that the generated user feature is considered to be invalid once the user feature exceeds the effective time window. Therefore, the problem that once one user feature is extracted, the user feature is directly used as a permanent user feature to cause inaccuracy is avoided, and meanwhile, the problem that in some cases, the user feature is extracted inaccurateness due to the permanent user feature directly caused by user operation can be reduced through the processing of converting short-term user features into long-term user features.

In some embodiments, the parsing unit 120 is specifically configured to generate a first validity time window of the short-term user feature and a second validity time window of the long-term user feature according to the duration of the operation time window; wherein the duration of the second validity time window is greater than the duration of the first validity time window.

In this embodiment, the analysis unit 120 may extract the long-term user features and the short-term user features, and may consider both the short-term interest and the long-term user interest of the user, so that the extracted user features may indicate the interest of the user as much as possible, and accuracy of the user features may be improved.

In some embodiments, the parsing unit 120 is further specifically configured to delete the first potential long-term feature from the set of potential long-term features if the reverse feature of the first potential long-term feature is a first type of official user feature in the set of official user features; the first type of formal user characteristics are formal user characteristics with the ranking value in a preset range; the scoring value is used for controlling the recommendation probability of the information corresponding to the formal user characteristics.

By intersection processing with formal user features, user features formed by misoperation can be reduced, and the accuracy of user feature extraction is improved again.

In some embodiments, the parsing unit 120 is further specifically configured to adjust the score value of the reverse feature of the first potential long-term feature if the first potential long-term feature is a second type of formal user feature in the set of formal user features; the second type of formal user characteristics are characteristics of which the score value is outside the predetermined range.

By adjusting the score value, the recommendation probability is dynamically adjusted by adjusting the score value in the process that the user has reverse interest of a certain formal user characteristic, so that the current requirement or interest of the user is met as much as possible.

In some embodiments, the parsing unit 120 is specifically configured to count the occurrence frequency of the short-term user feature within the second operation time window; when the frequency of occurrence is greater than a predetermined threshold, converting the short-term user characteristics to the long-term user characteristics.

In the embodiment, the short-term user characteristics and the long-term user characteristics are used for conversion, so that the data processing amount is reduced, and the data processing complexity is reduced.

In some embodiments, the parsing unit 120 is further configured to parse negative-going operational behavior data within the first operational time window to generate a short-term negative-going feature; analyzing negative operation behavior data in the second operation time window to generate long-term negative characteristics; wherein the negative-going operational behavior characterized by the negative-going operational behavior data is an operation that reduces display of information.

For example, the negative-going operational behavior includes: a delete operation and/or a first type close operation; the first type of closing operation comprises closing operation of information links and/or closing operation of which detailed page display time is less than a first preset time length; the information link is: and entering a link of a detailed page of the detailed information display corresponding to the information link.

The negative operation behavior data and the negative operation behavior may be referred to in the foregoing embodiments, and are not repeated here.

Several specific examples are provided below in connection with the above embodiments:

example one:

the example is divided according to dislike behaviors input by a user, mainly divided into short-term behaviors and long-term behaviors, and different strategies are adopted for the two types of behaviors to obtain short-term user characteristics and long-term user characteristics. And in the short-term behavior expression, all the characteristics on the behavior are pressed within a certain time window, and when the time window is exceeded, the characteristics are tried to be exposed again, and the continuous expression result of the user is observed to obtain the short-term user characteristics. And the long-term behavior is obtained by comprehensively judging the aversion degree of the user to a certain label/classification through accumulating the historical dislike occurrence frequency of a certain characteristic and comparing the portrait characteristics of the forward behavior of the user. In the example, on one hand, the active behavior expression of the user is respected, and meanwhile, the historical behavior and the forward image feature of the user are considered, so that the misjudgment caused in the filtering scheme aiming at the single user behavior is reduced to the greatest extent. The forward image feature is a set of the forward features.

Example two:

as shown in fig. 7, the present example provides an information system architecture comprising:

a Business and Operation Support System (BOSS) reporting System, a database (Data Base, DB), a dislike Data indexing System, an article/video mapping System, an article tagging System, a video tagging System, and a user imaging System.

And the BOSS reporting system performs real-time subscription analysis processing, and sends the analyzed user operation behavior data to the database and the dislike data index system for index processing of the dislike data index system, so that short-term dislike feature extraction and long-term dislike feature extraction are performed to obtain the dislike features. The dislike feature may be used for user portrayal and/or storage of the dislike feature.

The article/video mapping system is mainly used for extracting the video index from the article index.

The article label system can be mainly used for extracting the image-text article labels. The article tag here corresponds to an information tag. The information tag is an information feature and is used for describing the information type or the information content of certain information. The video tag system is used for providing tags of video articles. The user portrait system mainly stores original user portrait characteristics before the current time. The original portrait features here are a user feature set.

Dislike feature extraction strategy:

the data source is as follows: the user operation behavior data is extracted and reported from the Boss system by the Boss reporting system, the flow data of the user is obtained through real-time subscription, the analyzed data is stored in the dislike data index system, and meanwhile, the data is backed up and written into the database DB. The running water data here is data of various running water records generated by the schedule operation.

Triggering a calculation opportunity:

and triggering and calculating the dislike portrait characteristics of the user every time a new dislike behavior of the user is received, so that the expression of the user is fed back quickly.

Periodically acquiring a user list with dislike behaviors in the DB for a period of time, and triggering and calculating the corresponding dislike portrait characteristics, so that the short-term dislike characteristics of the user are released, and the change caused by the change of the basic portrait characteristics in the historical characteristic calculation is solved;

and outputting a result: the method comprises two parts, wherein one part is label/classification characteristics which are definitely judged to be disliked, and the characteristics are subjected to strict striking filtering in subsequent recommendation; one part is a label/classification judged to be possibly disliked, and the original user portrait needs to be subjected to weight reduction, and the appearance probability of the class of characteristics is reduced in subsequent recommendation;

as shown in fig. 8, the user feature extraction may include:

step 1: the obtaining of the index of articles disliked by the user may specifically include obtaining all article lists indexed by the user, and finding an index of an article that clearly indicates the disliked by the user through the user operation from all the article lists. The articles herein may include various forms of messages, such as, for example, teletext messages and video messages. The image-text message is a pure text message which only comprises words, can also be a pure picture message which only comprises pictures, and can also comprise an image-text mixed message which simultaneously comprises words and pictures. The video message may be a message including video.

Step 2: the original data is divided into a teletext article and a video article through the mapping from the article to the video Identification (ID). The text-text articles are the aforementioned text-text messages, and the video-text articles are the video messages.

And step 3: respectively acquiring label/classification information of the image-text articles and label/classification information of the video articles;

and 4, step 4: and judging the occurrence time of the dislike operation, wherein the short-term user feature extraction process and the long-term user feature extraction process are divided according to the dislike operation time by taking the last 24 hours as a boundary. The dislike operation is one of the above-mentioned reverse operations.

And 5: short-term user feature extraction:

the aggregate label/classification, for example, may include: the labels/classifications of the graphics-text articles and the video articles are all determined as characteristics that the user definitely dislikes.

Step 6: extracting long-term user features:

step 6.1: aggregating labels/classifications, counting times; can include the following steps: and aggregating the labels/classifications of the image-text articles and the video articles to obtain the occurrence frequency of each label/classification.

Step 6.2: acquiring an original portrait of a user (by feature extraction through forward recessive behaviors such as playing), and taking intersection with a current disliked label/classification set, wherein the intersection shows that the classification/label has forward behavior features, and executing step 6.3; the step 6.4 is executed outside the intersection.

Step 6.3: when the mark of the label/classification on the original portrait of the user is higher than the mark threshold of the top 20, the label/classification which is really liked by the user is considered, and the dislike characteristic is ignored; conversely, for tags/classifications below the top 20 score threshold, the original portrait score is de-weighted according to the number of cumulative occurrences, reducing its probability of occurrence in the recommendation. The score here is one of the aforementioned scores.

Step 6.4: comparing the occurrence frequency of the labels/classifications, and when the occurrence frequency is larger than a specified threshold (each label/classification can specify different thresholds according to the respective characteristics), determining that the labels/classifications are definitely dislike characteristics; otherwise, the accumulation is considered to be insufficient, and the judgment output is temporarily not made.

And 7: and combining the short-term dislike feature and the long-term dislike feature extraction result, and outputting a user-defined dislike feature list and a feature list needing to perform score reduction on the original portrait. The short-term dislike feature here is the aforementioned short-term negative-going feature. The long-term dislike feature is the aforementioned long-term negative-going feature.

Based on the scheme for extracting the behavior features which are not interesting to the user, the beneficial effects brought by the scheme comprise the following steps:

(1) through historical data accumulation and conversion analysis of short-term behaviors of the user to long-term behaviors, characteristics which are not interesting to the user are more accurately analyzed and extracted, more accurate user portrait basis is provided for subsequent recommendation, and content loss caused by portrait accidental injury is avoided to the greater extent;

(2) when the user portrait features are more accurate, the overall click rate and the playing completion degree of the recommended content can be well improved.

Example three:

as shown in fig. 9, this example provides an alternative hardware structure diagram of a user feature extraction apparatus, which includes a processor 11, an input/output interface 13 (e.g., a display screen, a touch screen, and a speaker), a storage medium 14, and a network interface 12, and the components may be connected to communicate via a system bus 15. Accordingly, the storage medium 14 stores therein executable instructions for executing the service processing method according to the embodiment of the present invention. The hardware modules shown in fig. 9 may be partially implemented, fully implemented, or other hardware modules as needed, the number of each type of hardware module may be one or more, and the hardware modules may be implemented in the same geographical location, or distributed in different geographical locations, and may be used to perform the user feature extraction methods shown in fig. 1, fig. 5, and fig. 8.

In the several embodiments provided in the present application, it should be understood that the disclosed apparatus and method may be implemented in other ways. The above-described device embodiments are merely illustrative, for example, the division of the unit is only a logical functional division, and there may be other division ways in actual implementation, such as: multiple units or components may be combined, or may be integrated into another system, or some features may be omitted, or not implemented. In addition, the coupling, direct coupling or communication connection between the components shown or discussed may be through some interfaces, and the indirect coupling or communication connection between the devices or units may be electrical, mechanical or other forms.

The units described as separate parts may or may not be physically separate, and parts displayed as units may or may not be physical units, that is, may be located in one place, or may be distributed on a plurality of network units; some or all of the units can be selected according to actual needs to achieve the purpose of the solution of the embodiment.

In addition, all the functional units in the embodiments of the present invention may be integrated into one processing module, or each unit may be separately used as one unit, or two or more units may be integrated into one unit; the integrated unit can be realized in a form of hardware, or in a form of hardware plus a software functional unit.

Those of ordinary skill in the art will understand that: all or part of the steps for implementing the method embodiments may be implemented by hardware related to program instructions, and the program may be stored in a computer readable storage medium, and when executed, the program performs the steps including the method embodiments; and the aforementioned storage medium includes: a mobile storage device, a Read-Only Memory (ROM), a Random Access Memory (RAM), a magnetic disk or an optical disk, and other various media capable of storing program codes.

The above description is only for the specific embodiments of the present invention, but the scope of the present invention is not limited thereto, and any person skilled in the art can easily conceive of the changes or substitutions within the technical scope of the present invention, and all the changes or substitutions should be covered within the scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the appended claims.

Claims

1. A user feature extraction method is characterized by comprising the following steps:

acquiring user operation behavior data;

analyzing the operation behavior data in the first operation time window, and generating short-term user characteristics according to the occurrence frequency of each operation behavior;

converting the short-term user features into long-term user features when the occurrence frequency of the short-term user features in a second operation time window is greater than a preset threshold value; wherein the duration of the second operating time window is greater than the duration of the first operating time window; determining a first effective time window of the short-term user characteristics according to the first operation time window, and determining a second effective time window of the long-term user characteristics according to the second operation time window; wherein the duration of the second validity time window is greater than the duration of the first validity time window;

determining that the short-term user characteristics are invalid when the time of the first effective time window is reached, and determining that the long-term user characteristics are invalid when the time of the second effective time window is reached; the short-term user characteristics are used for controlling the recommendation probability of the first type of information in the first effective time window, and the long-term user characteristics are used for controlling the recommendation probability of the second type of information in the second effective time window.

2. The method of claim 1, further comprising:

3. The method of claim 2, further comprising:

if the reverse feature of the first potential long-term feature is a first type of official user feature in the official user feature set, deleting the first potential long-term feature from the potential long-term feature set; the first type of formal user characteristics are formal user characteristics with the ranking value in a preset range; the scoring value is used for controlling the recommendation probability of the information corresponding to the formal user characteristics.

4. The method of claim 3, further comprising:

if the first potential long-term feature is a second type of formal user feature in the formal user feature set, adjusting the score value of the reverse feature of the first potential long-term feature; the second type of formal user characteristics are characteristics of which the score value is outside the predetermined range.

5. The method of claim 1, further comprising:

6. The method of claim 5,

the negative operation comprises: a delete operation and/or a first type close operation; the first type of closing operation comprises closing operation of information links and/or closing operation of which detailed page display time is less than a first preset time length; the information link is: and entering a link of a detailed page of the detailed information display corresponding to the information link.

7. A user feature extraction device, characterized by comprising:

the acquisition unit is used for acquiring user operation behavior data;

the analysis unit is used for analyzing the operation behavior data in the first operation time window, generating short-term user characteristics according to the occurrence frequency of each operation behavior, and converting the short-term user characteristics into long-term user characteristics when the occurrence frequency of the short-term user characteristics in the second operation time window is greater than a preset threshold value; wherein the duration of the second operating time window is greater than the duration of the first operating time window;

a determining unit, configured to determine a first effective time window of the short-term user characteristics according to the first operation time window, and determine a second effective time window of the long-term user characteristics according to the second operation time window; wherein the duration of the second validity time window is greater than the duration of the first validity time window;

the determining unit is further used for determining that the short-term user characteristics are invalid when the time of the first effective time window arrives, and determining that the long-term user characteristics are invalid when the time of the second effective time window arrives;

the short-term user characteristics are used for controlling the recommendation probability of the first type of information in the first effective time window, and the long-term user characteristics are used for controlling the recommendation probability of the second type of information in the second effective time window.

8. The apparatus of claim 7,

the analysis unit is specifically configured to analyze the operation behavior data to obtain the occurrence frequency of the potential long-term features in the potential long-term feature set; determining whether a reverse feature corresponding to a first potential long-term feature included in the potential long-term feature set is a formal user feature in the formal user feature set; and if the reverse characteristic of the first potential long-term characteristic is not the formal user characteristic, determining that the first potential characteristic is the long-term user characteristic according to the occurrence frequency.

9. The apparatus of claim 8,

the analysis unit is further specifically configured to delete the first potential long-term feature from the potential long-term feature set if the reverse feature of the first potential long-term feature is a first type of official user feature in the official user feature set; the first type of formal user characteristics are formal user characteristics with the ranking value in a preset range; the scoring value is used for controlling the recommendation probability of the information corresponding to the formal user characteristics.

10. The apparatus of claim 9,

the analysis unit is further specifically configured to adjust a score value of a reverse feature of the first potential long-term feature if the first potential long-term feature is a second type of official user feature in the official user feature set; the second type of formal user characteristics are characteristics of which the score value is outside the predetermined range.

11. The apparatus of claim 7,

the analysis unit is further configured to analyze negative operation behavior data in the first operation time window to generate a short-term negative feature; analyzing negative operation behavior data in the second operation time window to generate long-term negative characteristics; wherein the negative-going operational behavior characterized by the negative-going operational behavior data is an operation that reduces display of information.

12. The apparatus of claim 11,

the negative-going operational behavior comprises: a delete operation and/or a first type close operation; the first type of closing operation comprises closing operation of information links and/or closing operation of which detailed page display time is less than a first preset time length; the information link is: and entering a link of a detailed page of the detailed information display corresponding to the information link.

13. A storage medium having stored thereon a computer program, wherein the computer program, when executed by a processor, implements the user feature extraction method of any one of claims 1-6.