Method and apparatus for generating interest feature of user, and electronic device

By generating user interest characteristics and behavioral object characteristics, combining clustering and prediction models, the problem of insufficient recommendation effects caused by the sparseness of long-tail user data in e-commerce platforms is solved, and the accuracy and personalization of the recommendation system are improved.

WO2024230283A1PCT designated stage expired Publication Date: 2024-11-14HANGZHOU ALIBABA INT INTERNET IND CO LTD

Patent Information

Application Number
PCT/CN2024/079456
Authority / Receiving Office
WO · WO
Patent Type
Applications
Current Assignee / Owner
Priority Date
2023-05-05
Filing Date
2024-02-29
Publication Date
2024-11-14

AI Technical Summary

Technical Problem

The behavioral data of long-tail users in e-commerce platforms is sparse, which leads to insufficient accuracy and personalization of recommendation systems for their products, and is unable to effectively improve user stickiness and recommendation effects.

Method used

By generating user interest characteristics, the behavior object feature vectors of user behavior objects are obtained, the correlation between positive and negative feedback objects is calculated, the initial interest feature expression is generated, and the accuracy of the recommendation system is improved through clustering and prediction models. .

Benefits of technology

It improves the accuracy and user stickiness of product recommendations for long-tail users, enhances the personalization and applicability of the recommendation system, and solves the problem of insufficient recommendation effects caused by data sparseness.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN2024079456_14112024_PF_FP_ABST
    Figure CN2024079456_14112024_PF_FP_ABST
Patent Text Reader

Abstract

Disclosed in the present invention is a method for generating an interest feature of a user. The method comprises: acquiring behavior object feature vectors which correspond to a target time window and are used for representing behavior objects of a user, wherein the behavior object feature vectors comprise positive-feedback object feature vectors and negative-feedback object feature vectors; performing integrated calculation on the negative-feedback object feature vectors, so as to obtain an integrated negative-feedback object vector; calculating the correlation between the integrated negative-feedback object vector and each positive-feedback object feature vector; and according to the correlation, the positive-feedback object feature vectors and the integrated negative-feedback object vector, generating a primary interest feature expression of the user, which primary interest feature expression corresponds to the target time window. By using the method, an interest feature of a user can be accurately described, thus providing a foundation for accurate recommendation. The present invention further relates to a user interest feature clustering method, a user behavior estimation method and a method for training a user behavior object estimation model; and the present invention also provides apparatuses related to the methods, and an electronic device and a storage medium.
Need to check novelty before this filing date? Find Prior Art

Description

Method, device and electronic device for generating user interest features

[0001] This application claims priority to the Chinese patent application filed with the China Patent Office on May 5, 2023, with application number 202310508748.1 and application name “A method, device and electronic device for generating user interest features”, the entire contents of which are incorporated by reference into this application. Technical Field

[0002] The present invention relates to the field of e-commerce technology, and more particularly to a method for generating user interest features. The present invention also relates to a method for clustering user interest features, a method for predicting user behavior, and a method for training a user behavior object prediction model. The present application also relates to devices, electronic devices, and storage media related to each of the aforementioned methods. Background Art

[0003] With the development of e-commerce technology, precision marketing is widely used on major e-commerce platforms. As a core tool in precision marketing, recommendation systems can recommend products and information of interest to users when faced with a vast amount of product information.

[0004] In the current precision marketing model, recommendation systems have significant flaws in their recommendations for the long-tail market. Specifically, a small number of top users have a high frequency of activity on e-commerce platforms and a large number of purchase interaction records. Therefore, recommendation systems can use this extensive behavioral data to provide them with highly accurate, personalized, and applicable products and information. However, the majority of long-tail users have a low frequency of activity on the platform and relatively few purchase interaction records. Therefore, the behavioral data available to the recommendation system is relatively sparse, and the effectiveness of recommendations for long-tail users is significantly weaker than for top users.

[0005] To address the problem of sparsity in behavioral data for recommendations, the current general solution is to import auxiliary information other than user behavior into the recommendation system, such as user profiles and product texts, and incorporate them into the recommendation algorithm together with user behavior data. However, this auxiliary information is not always available. For example, with increasingly stringent restrictions on user privacy terms, obtaining user profiles has become increasingly difficult. The extraction of product texts is highly dependent on the maturity of NLP (Natural Language Processing) technology. When faced with texts in minority languages, NLP technology still faces many challenges.

[0006] Therefore, for the huge number of long-tail users on e-commerce platforms, how to improve the accuracy of product recommendations and user click-through rate estimation in the recommendation system without the use of auxiliary information has become an urgent problem to be solved.

[0007] Summary of the Invention

[0008] In order to overcome the deficiencies of the prior art, one of the objectives of the present invention is to provide a method for generating user interest features so as to accurately describe the user's interest features and thereby provide a basis for more accurate recommendations.

[0009] Another object of the present invention is to provide a method for clustering user interest features, which uses the interest features obtained by the above method and realizes accurate clustering of user interest features, thereby providing the possibility for more accurate recommendations.

[0010] A further object of the present invention is to provide a user behavior prediction method, which uses the above-mentioned clustering results of user interest characteristics and the obtained user interest characteristics to predict the target user's behavior possibility towards the recommended object for the target user and the specific recommended object.

[0011] In a first aspect, the present application provides a method for generating user interest features, comprising: obtaining a behavior object feature vector corresponding to a target time window for characterizing a user's behavior object; the behavior object feature vector for characterizing the user's behavior object comprises a positive feedback object feature vector and a negative feedback object feature vector; integrating and calculating the negative feedback object feature vector to obtain a negative feedback object integrated vector corresponding to the negative feedback object feature vector; calculating the correlation between the negative feedback object integrated vector and each of the positive feedback object feature vectors; and generating a primary interest feature expression of the user corresponding to the target time window based on the correlation, the positive feedback object feature vector, and the negative feedback object integrated vector.

[0012] Optionally, obtaining a behavior object feature vector corresponding to a target time window for characterizing a user's behavior object includes: obtaining a user behavior sequence sample; grouping the user behavior sequence samples corresponding to preset time windows to obtain user behavior group samples corresponding to each time window; obtaining a behavior object corresponding to each user behavior group sample; and generating the behavior object feature vector corresponding to the target time window within a selected target time window based on the behavior object corresponding to the user behavior.

[0013] Optionally, within the selected target time window, based on the behavior object corresponding to the user behavior, the behavior object feature vector corresponding to the target time window is generated, including: according to preset filtering conditions, the behavior objects of the users in the selected target time window are filtered to obtain sample behavior objects; based on the sample behavior objects, the behavior object feature vector corresponding to the target time window is generated.

[0014] Optionally, the preset screening conditions include one or more of the following factors: the frequency of the user's behavior on the behavior object reaches a preset value, the duration of the user's behavior on the behavior object reaches a preset value, and the type of the user's behavior on the behavior object meets the preset type.

[0015] Optionally, generating the user's primary interest feature expression corresponding to the target time window based on the correlation, the positive feedback object feature vector, and the negative feedback object integrated vector includes: obtaining a correlation index between each of the positive feedback object feature vectors and the negative feedback object integrated vector; taking the positive feedback object feature vector whose correlation index is lower than a preset correlation threshold as an available positive feedback object feature vector; and generating the user's primary interest feature expression based on the correlation, the available positive feedback object feature vector, and the negative feedback object integrated vector.

[0016] Optionally, generating the user's primary interest feature expression based on the correlation, the available positive feedback object feature vector and the negative feedback object integrated vector includes: assigning a second weight to each of the available positive feedback object feature vectors in a preset manner based on the correlation between the available positive feedback object feature vectors and the negative feedback object integrated vector; wherein, the higher the correlation, the lower the weight; and using the negative feedback object integrated vector and the weighted positive feedback object feature vector to generate the user's primary interest feature expression.

[0017] Optionally, the use of the negative feedback object integrated vector and the weighted positive feedback object feature vector to generate the user's primary interest feature expression includes: inputting each weighted positive feedback object feature vector into a first transformation neural network to obtain a positive feedback object integrated vector; and fusing the positive feedback object integrated vector with the negative feedback object integrated vector to obtain the user's primary interest feature expression.

[0018] Optionally, before the step of generating the user's primary interest feature expression corresponding to the target time window based on the correlation, the positive feedback object feature vector, and the negative feedback object integrated vector, the following steps are included: assigning a first weight to the corresponding positive feedback object feature vector based on the frequency of each positive feedback object as a user behavior object; in the step of generating the user's primary interest feature expression corresponding to the target time window based on the correlation, the positive feedback object feature vector, and the negative feedback object integrated vector, the positive feedback object feature vector assigned the first weight is used.

[0019] Optionally, the fused positive feedback object integrated vector and the negative feedback object integrated vector are spliced.

[0020] The second aspect of the present application provides a method for clustering user interest features, the method comprising: obtaining a primary interest feature expression and an interest category feature expression of a target user; the interest category feature expression is a feature expression of each interest category formed by classifying users according to their interests; mapping and fusion processing is performed on the primary interest feature expression of the target user according to the interest category feature expression to obtain the interest feature expression code of the target user; encoding the interest feature expression of the target user and decoding the encoding to obtain the decoded interest feature expression code; calculating the loss difference between the decoded interest feature expression code and the preset target interest feature expression, and adjusting the encoding model, decoding model, and interest category feature expression according to the loss difference until the loss difference reaches a predetermined target; and obtaining the interest category expression to which the target user belongs under the target time window according to the dimensionality reduction feature expression at this time.

[0021] Optionally, the target interest feature expression adopts the primary interest feature expression of the user, or adopts a high-order interest feature expression; the high-order interest feature expression is obtained by associating the primary interest feature expression of the target time window with the primary interest feature expressions of other time windows.

[0022] According to a third aspect of the present application, a method for estimating user behavior objects is provided, comprising: obtaining user behavior samples divided according to time windows; generating a behavior object feature vector for characterizing the user's behavior object based on the user behavior samples; the behavior object feature vector for characterizing the user's behavior object includes a positive feedback object feature vector and a negative feedback object feature vector; generating a primary interest feature expression corresponding to each time window of the user based on the behavior object feature vector; interactively processing the primary interest feature expression of each time window with the primary interest feature expression of other time windows to obtain a high-order interest feature expression for each time window; using the primary interest feature expression and the high-order interest feature expression of each time window to classify users into types according to their interests, obtaining a time window interest category expression to which a target user belongs in each time window; fusing the time window target interest category expression to which the target user belongs in each time window and the user's high-order interest feature expression corresponding to each time window to obtain a target interest category feature expression for the target user; and using the target interest category expression of the target user to estimate the probability of the target user's behavior towards the target behavior object.

[0023] Optionally, based on the behavior object feature vector, the user's primary interest feature expression corresponding to each time window is generated, wherein the process of generating the primary interest feature expression corresponding to the target time window includes: obtaining the behavior object feature vector corresponding to the target time window for characterizing the user's behavior object; the behavior object feature vector for characterizing the user's behavior object includes a positive feedback object feature vector and a negative feedback object feature vector; integrating the negative feedback object feature vector to obtain a negative feedback object integration vector corresponding to the negative feedback object feature vector; calculating the correlation between the negative feedback object integration vector and each of the positive feedback object feature vectors; generating the user's primary interest feature expression corresponding to the target time window based on the correlation, the positive feedback object feature vector, and the negative feedback object integration vector.

[0024] Optionally, the primary interest feature expression and high-order interest feature expression of each time window are used to classify users into types according to their interests to form interest category feature expressions of each interest category, and obtain the time window interest category expression to which the target user belongs in each time window, including: obtaining the primary interest feature expression and interest category feature expression of the target user in the target time window; mapping and fusing the primary interest feature expression of the target user according to the interest category feature expression to obtain the interest feature expression code of the target user; encoding the interest feature expression of the target user and decoding it to obtain the decoded interest feature expression code; calculating the loss difference between the decoded interest feature expression code and the high-order interest feature expression of the user in the target time window, and adjusting the encoding model, decoding model, and interest category feature expression according to the loss difference until the loss difference reaches a predetermined target; and obtaining the interest category expression to which the target user belongs in the target time window according to the dimensionality reduction feature expression at this time.

[0025] Optionally, the use of the target interest category expression of the target user to estimate the behavior probability of the target user towards the target behavior object includes: performing attention mechanism processing on the target interest category expression of the target user and the characteristic expression of the target behavior object to obtain the attention mechanism expression of the target user; providing the attention mechanism expression of the target user and the characteristic expression of the target behavior object to a pre-trained machine learning model to obtain the behavior probability of the target user towards the target behavior object.

[0026] The fourth aspect of the present application provides a training method for a user behavior object prediction model, comprising: obtaining user behavior samples divided according to time windows, including behavior objects involved in user behavior; corresponding to each time window, generating a behavior object feature vector for characterizing the user's behavior object in the input layer of the existing model; the behavior object feature vector for characterizing the user's behavior object includes a positive feedback object feature vector and a negative feedback object feature vector; based on the behavior object feature vector, generating the user's primary interest feature expression corresponding to each time window in the user interest extraction model; interactively processing the primary interest feature expression of each time window with the primary interest feature expression of other time windows through a transformation neural network to obtain a high-order interest feature expression corresponding to each time window; using the primary interest feature expression, high-order interest feature expression and other time windows to extract the user's behavior object; extracting the user's behavior object from the model ... Expression, as well as the interest category feature expression of each interest category formed by dividing users into types according to their interests, in the user interest clustering model, the time window interest category expression to which the target user belongs corresponding to each time window is obtained; the time window target interest category expression to which the target user belongs under the said each time window and the user's high-order interest feature expression corresponding to each time window are fused to obtain the target interest category expression of the target user; the target interest category expression of the target user is used to estimate the behavior probability of the target user for each behavior object involved in the user behavior sample; the estimated result is compared with the actual record of the user's behavior for the behavior object in the user behavior sample, and according to the comparison result, the parameters of each model involved in the user behavior object prediction model are adjusted until the predetermined standard is met.

[0027] The fifth aspect of the present application provides a method for generating user interest features, which is applied to a cloud service end, and the method includes: obtaining an interest feature request message sent by a client to request obtaining user interest features; returning the page content associated with the user interest features to the client in response to the interest feature request message; wherein the page content associated with the user interest features is obtained based on the estimated behavior probability of the target user to the target behavior object; the estimated behavior probability of the target user to the target behavior object is obtained in the following manner: obtaining user behavior samples divided according to time windows; generating a behavior object feature vector for characterizing the user's behavior object based on the user behavior samples; the behavior object feature vector for characterizing the user's behavior object includes a positive feedback object feature vector and a negative feedback object feature vector; generating the The user's primary interest feature expression corresponding to each time window; interactively processing the primary interest feature expression of each time window with the primary interest feature expression of other time windows to obtain the high-order interest feature expression of each time window; using the primary interest feature expression and high-order interest feature expression of each time window, and dividing the user into types according to the user's interests to form the interest category feature expression of each interest category, to obtain the time window interest category expression to which the target user belongs under each time window; fusing the time window target interest category expression to which the target user belongs under each time window and the user's high-order interest feature expression corresponding to each time window to obtain the target interest category feature expression of the target user; using the target interest category expression of the target user to estimate the behavior probability of the target user towards the target behavior object.

[0028] The sixth aspect of the present application provides a method for generating user interest features, which is applied to a client, and the method includes: sending an interest feature request message to a cloud server for requesting to obtain user interest features; receiving page content associated with the user interest features returned by the cloud server; in response to detecting a trigger operation of the target user on a preset page content, displaying the page content associated with the user interest features; or, in response to detecting a trigger operation of the target user on a query object, displaying the page content associated with the user interest features; wherein the page content associated with the user interest features is obtained by the cloud server according to the estimated behavior probability of the target user on the target behavior object according to the method provided in the fifth aspect above.

[0029] In a seventh aspect, the present application provides a device for generating user interest features, comprising: an acquisition unit for acquiring a behavior object feature vector corresponding to a target time window for characterizing a user's behavior object; the behavior object feature vector for characterizing the user's behavior object comprises a positive feedback object feature vector and a negative feedback object feature vector; a processing unit for integrating and calculating the negative feedback object feature vector to obtain a negative feedback object integrated vector corresponding to the negative feedback object feature vector; a calculation unit for calculating the correlation between the negative feedback object integrated vector and each of the positive feedback object feature vectors; and an interest feature expression generation unit for generating a primary interest feature expression of the user corresponding to the target time window based on the correlation, the positive feedback object feature vector, and the negative feedback object integrated vector.

[0030] In an eighth aspect, the present application provides a device for clustering user interest features, comprising: an acquisition unit for acquiring a target user's primary interest feature expression and an interest category feature expression; the interest category feature expression is a feature expression of each interest category formed by classifying users according to their interests; a processing unit for mapping and fusing the target user's primary interest feature expression according to the interest category feature expression to obtain the target user's interest feature expression code; a construction unit for encoding the target user's interest feature expression, performing decoding processing, and obtaining a decoded interest feature expression code; a calculation unit for calculating the loss difference between the decoded interest feature expression code and a preset target interest feature expression, and adjusting the encoding model, decoding model, and interest category feature expression according to the loss difference until the loss difference reaches a predetermined target; and obtaining the interest category expression to which the target user belongs under the target time window according to the dimensionality reduction feature expression at this time.

[0031] The ninth aspect of the present application provides a device for estimating user behavior objects, comprising: an acquisition unit for acquiring user behavior samples divided according to time windows; a first processing unit for generating a behavior object feature vector for characterizing the user's behavior object based on the user behavior samples; the behavior object feature vector for characterizing the user's behavior object includes a positive feedback object feature vector and a negative feedback object feature vector; a second processing unit for generating a primary interest feature expression of the user corresponding to each time window based on the behavior object feature vector; a feature expression acquisition unit for interactively processing the primary interest feature expression of each time window with the primary interest feature expression of other time windows to obtain a high-order interest feature expression for each time window. An interest category division unit is configured to use the primary interest feature expression and high-order interest feature expression of each time window to divide users into types according to their interests to form interest category feature expressions of each interest category, and obtain the time window interest category expression to which the target user belongs under each time window; a fusion processing unit is configured to fuse the time window target interest category expression to which the target user belongs under each time window and the user's high-order interest feature expression corresponding to each time window, and obtain the target interest category feature expression of the target user; a behavior probability acquisition unit is configured to use the target interest category expression of the target user to estimate the behavior probability of the target user towards the target behavior object.

[0032] The tenth aspect of the present application provides a training device for a user behavior object prediction model, comprising: an acquisition unit for acquiring user behavior samples divided according to time windows, including behavior objects involved in user behavior; a first processing unit for generating, corresponding to each time window, a behavior object feature vector for characterizing the user's behavior object in the input layer of an existing model; the behavior object feature vector for characterizing the user's behavior object includes a positive feedback object feature vector and a negative feedback object feature vector; a feature expression generation unit for generating, in a user interest extraction model, a primary interest feature expression of the user corresponding to each time window based on the behavior object feature vector; an interactive processing unit for interactively processing the primary interest feature expression of each time window with the primary interest feature expression of other time windows through a transformation neural network to obtain a high-order interest feature expression corresponding to each time window; a category expression acquisition unit for using the primary interest feature expression of each time window to extract the high-order interest feature expression. Expression, high-order interest feature expression, and interest category feature expression of each interest category formed by dividing users into types according to user interests, in the user interest clustering model, the time window interest category expression to which the target user belongs corresponding to each time window is obtained; a fusion processing unit is used to fuse the time window target interest category expression to which the target user belongs under the said each time window, and the user's high-order interest feature expression corresponding to each time window, to obtain the target interest category expression of the target user; a behavior probability estimation unit is used to use the target interest category expression of the target user to estimate the estimated result of the behavior probability of the target user for each behavior object involved in the user behavior sample; a parameter adjustment unit is used to compare the estimated result with the actual record of the user's behavior for the behavior object in the user behavior sample, and according to the comparison result, adjust the parameters of each model involved in the user behavior object prediction model until the predetermined standard is met.

[0033] In an eleventh aspect of the present application, there is provided a device for generating user interest features, which is applied to a cloud service end, and comprises: an acquisition unit for obtaining an interest feature request message issued by a client for requesting to obtain user interest features; a transmission unit for returning the page content associated with the user interest features to the client in response to the interest feature request message; wherein, the page content associated with the user interest features is obtained based on the estimated behavior probability of the target user to the target behavior object; the estimated behavior probability of the target user to the target behavior object is obtained in the following manner: obtaining user behavior samples divided according to time windows; generating a behavior object feature vector for characterizing the user's behavior object based on the user behavior samples; the behavior object feature vector for characterizing the user's behavior object includes a positive feedback object feature vector and a negative feedback object feature vector; and generating a behavior object feature vector based on the behavior object feature vector. to generate the primary interest feature expressions of the user corresponding to each time window; interactively process the primary interest feature expressions of each time window with the primary interest feature expressions of other time windows to obtain the high-order interest feature expressions of each time window; use the primary interest feature expressions and high-order interest feature expressions of each time window to classify users into types according to their interests to form interest category feature expressions of each interest category, and obtain the time window interest category expression to which the target user belongs under each time window; fuse the time window target interest category expression to which the target user belongs under each time window and the user's high-order interest feature expression corresponding to each time window to obtain the target interest category feature expression of the target user; use the target interest category expression of the target user to estimate the behavior probability of the target user towards the target behavior object.

[0034] The twelfth aspect of the present application provides a device for generating user interest features, which is applied to a client and includes: a sending unit for sending an interest feature request message to a cloud server for requesting to obtain user interest features; a receiving unit for receiving the page content associated with the user interest features returned by the cloud server; a first processing unit for displaying the page content associated with the user interest features in response to detecting a trigger operation of the target user on a preset page content; or, in response to detecting a trigger operation of the target user on a query object, displaying the page content associated with the user interest features; wherein the page content associated with the user interest features is obtained by the cloud server according to the target user's behavior probability towards the target behavior object estimated by the device provided in the above eleventh aspect.

[0035] The thirteenth aspect of the present application provides an electronic device comprising a processor and a memory; wherein the memory is used to store one or more computer instructions, wherein the one or more computer instructions are executed by the processor in accordance with any one of the above method embodiments.

[0036] In a fourteenth aspect, the present application provides a computer-readable storage medium having one or more computer instructions stored thereon, which are executed by a processor to implement any of the above method embodiments.

[0037] The method for generating user interest features provided by the present application includes: obtaining a behavior object feature vector corresponding to a target time window for characterizing a user's behavior object; the behavior object feature vector for characterizing the user's behavior object includes a positive feedback object feature vector and a negative feedback object feature vector; integrating and calculating the negative feedback object feature vector to obtain a negative feedback object integrated vector corresponding to the negative feedback object feature vector; calculating the correlation between the negative feedback object integrated vector and each of the positive feedback object feature vectors; and generating a preliminary interest feature expression of the user corresponding to the target time window based on the correlation, the positive feedback object feature vector, and the negative feedback object integrated vector. The preliminary interest feature expression provided by the above method can reflect the interest characteristics of the target user in the target time window, thereby providing initial data for the subsequent evaluation of the possibility of the user's behavior (such as clicking) on ​​the behavior object, such as a product.

[0038] In the method for clustering user interest features further provided in the present application, first, the target user's primary interest feature expression and interest category feature expression are obtained; the interest category feature expression is a feature expression of each interest category formed by dividing the user into types according to the user's interests; then, according to the interest category feature expression, the target user's primary interest feature expression is mapped and fused to obtain the target user's interest feature expression code; then, the target user's interest feature expression is encoded and decoded to obtain the decoded interest feature expression code; finally, the loss difference between the decoded interest feature expression code and the preset target interest feature expression is calculated, and the encoding model, decoding model, and interest category feature expression are adjusted according to the loss difference until the loss difference reaches the predetermined target; and according to the dimensionality reduction feature expression at this time, the interest category expression to which the target user belongs in the target time window is obtained. The above-mentioned method for clustering user interest features can provide the interest category expression to which the target user belongs, and use the interest category expression to deduce the behavior objects that the target user may be interested in, so that the various data collected are fully utilized to avoid the inability to make accurate recommendations due to too little data on the target user.

[0039] This application further provides a method for estimating user behavior targets. This method utilizes the primary user interest features obtained in the aforementioned steps, the higher-level user interest features derived from these primary user interest features, and the user interest category expressions obtained through user clustering. This information is then integrated to derive the user's target interest category expression. This target user's target interest category expression is then used to estimate the probability of the target user's behavior toward the target behavior target. Because this method uses the target user's target interest category as the basis for deriving their behavior, it can leverage the behavioral data of other similar customers, effectively solving the problem of accurately recommending to long-tail users and avoiding the problem of inaccurate recommendations due to overly sparse user data.

[0040] This application also provides a method for training a user behavior object prediction model, which can be used to implement end-to-end model training. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments recorded in this application. For ordinary technicians in this field, other drawings can also be obtained based on these drawings.

[0042] FIG1 is a schematic diagram of an application scenario of a method for generating user interest features in an embodiment of the present application;

[0043] FIG2 is a schematic diagram of a flow chart of the first embodiment of the present application;

[0044] FIG3 is a schematic diagram of a user behavior sequence sample corresponding to the first embodiment of the present application;

[0045] FIG4 is a diagram of the UBP model architecture corresponding to the first embodiment of the present application;

[0046] FIG5 is a flowchart of generating a first interest characteristic expression of a user according to the first embodiment of the present application;

[0047] FIG6 is a flow chart of a method for clustering user interest features according to the second embodiment of the present application;

[0048] FIG7 is a diagram of the UBC model architecture corresponding to the second embodiment of the present application;

[0049] FIG8 is a flowchart of a method for estimating user behavior objects according to the third embodiment of the present application;

[0050] FIG9 is a diagram of the HIM model architecture corresponding to the third embodiment of the present application;

[0051] FIG10 is a schematic diagram of feature vector transformation in the HIM model according to the third embodiment of the present application;

[0052] FIG11 is a flow chart of a method for training a user behavior object prediction model according to a fourth embodiment of the present application;

[0053] FIG12 is a flow chart of a method for generating user interest features according to a fifth embodiment of the present application;

[0054] FIG13 is a flow chart of a method for generating user interest features according to a sixth embodiment of the present application;

[0055] FIG14 is a schematic structural diagram of an apparatus for generating user interest features according to a seventh embodiment of the present application;

[0056] FIG15 is a schematic structural diagram of an apparatus for clustering user interest features according to an eighth embodiment of the present application;

[0057] FIG16 is a schematic structural diagram of a device for estimating user behavior objects according to a ninth embodiment of the present application;

[0058] FIG17 is a schematic structural diagram of a training device for a user behavior object prediction model according to a tenth embodiment of the present application;

[0059] FIG18 is a schematic structural diagram of an apparatus for generating user interest features according to the eleventh embodiment of the present application;

[0060] FIG19 is a schematic structural diagram of an apparatus for generating user interest features according to a twelfth embodiment of the present application;

[0061] FIG20 is a schematic diagram of the logical structure of the electronic device provided in the thirteenth embodiment of the present application. DETAILED DESCRIPTION

[0062] In order to make the purpose, advantages and features of the present application clearer, the following is a further detailed description of a method for generating user interest features, a method for clustering user interest features and a method for estimating user behavior objects proposed in the present application, in conjunction with the accompanying drawings and specific embodiments. In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention can be implemented in many other ways different from those described herein, and those skilled in the art can make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific implementation disclosed below.

[0063] It should be noted that, in the description of this application, the terms "first", "second", etc. are used for descriptive purposes only and are not to be understood as indicating or implying relative importance, or a specific order or precedence. For those skilled in the art, the specific meanings of the above terms in this application can be understood in specific circumstances. In addition, in the description of this application, unless otherwise specified, the term "plurality" refers to two or more. The term "and / or" describes the association relationship of associated objects and indicates that three relationships can exist. For example, A and / or B can represent: A exists alone, A and B exist at the same time, and B exists alone. The character " / " generally indicates that the associated objects before and after are in an "or" relationship. The terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or are inherent to these processes, methods, products or devices.

[0064] First, the terms involved in one or more embodiments of the present application are explained.

[0065] The click-through rate (CTR) of a product refers to the probability that a user clicks to view the details of a product when the product is displayed to the user. It is calculated as follows: click-through rate = clicks / exposures, where exposures refer to the number of times a product is displayed to the user, and clicks refer to the number of times a product is clicked by the user. The recommendation effect in the product recommendation system is measured by the estimated click-through rate of the product. Based on the estimated click-through rate of the product, the products recommended to the user are determined and sorted for display. By predicting the click-through rate of the user on the products in the candidate set, it is decided which product to display to the user. In this application, considering that users and other objects displayed on the network platform may also become click objects, and further considering that clicks are not the only way for users to associate with the objects displayed on the network platform, the "click" mentioned in the above explanation is referred to as behavior, and the "product" is referred to as behavior object; when describing the specific implementation method below, clicks or behaviors, as well as products or behavior objects, are selected according to the circumstances.

[0066] UBP Model: The User Behavior Pyramid (UBP) model, also known as a model for acquiring primary interest feature expressions, is a personalized interest extraction module based on confidence modeling of positive and negative feedback behavior objects, used to obtain personalized user interest expressions. The UBP model is a model specifically provided by this application to obtain primary user interest feature expressions based on usage scenarios, specifically implementing the method for generating user interest features provided in the first embodiment of this application.

[0067] UBC model: User Behavior Clustering (UBC), also known as the interest group feature expression acquisition model, is a user interest clustering module based on semi-supervised learning. It is used to obtain the user's semi-personalized interest expression. The final output result is to obtain the interest group representation of the target user, that is, to know which interest group the target user belongs to, and to give a mathematical expression of the interest group. Generally, each interest group is represented in the form of a feature vector corresponding to each interest group, which is referred to as the interest group representation of the target user. This model is used to execute the method of user interest feature clustering provided in the second embodiment of the present application.

[0068] HIM model: Hierarchical Interest Modeling (HIM), also known as the user behavior object prediction model, is a hierarchical interest modeling model based on the behavioral characteristics of users with few behaviors. Using this model, the probability of a user's behavior towards a target behavior object can be obtained based on the user's historical behavior data; in other words, the probability of a user clicking on a specific product can be obtained. The above-mentioned UBP model and UBC model are used in this model.

[0069] Next, in order to facilitate understanding of the various embodiments of the present application, the application background of the embodiments is described.

[0070] The number and variety of products on major e-commerce platforms are rapidly increasing. Users are faced with a vast amount of product information, requiring significant time and effort to search for desired products. Precision marketing, which can recommend products and information of interest to users, has become a popular choice across e-commerce platforms. As a core technical tool within this model, recommendation systems have become a research focus.

[0071] Traditional market trends adhere to the 80 / 20 rule (80% of sales come from 20% of popular brands). However, this rule is challenged when applied to internet-based e-commerce platforms. E-commerce platforms offer 99% of products the opportunity to be sold to users. Typically, popular products meet the needs of the majority of users, while less popular products can also represent the needs of smaller groups. Given the sheer volume of less popular products on e-commerce platforms, fully tapping into the market share of these less popular products requires thorough research into individual user needs and the ability to accurately recommend both popular and less popular products to those with the right needs. Currently, the precision marketing recommendation systems of major e-commerce platforms exhibit a significant "long tail" phenomenon. This section explains this phenomenon. The long tail refers to a small number of top users who perform a significant number of effective actions. These top users have a high frequency of activity on the platform and a rich record of purchase interactions. Based on this interaction data, recommendation systems can recommend products to them with high accuracy, strong personalization, and good adaptability. However, most users belong to the long-tail users in the long-tail phenomenon. They have a large user base, but their behavior frequency on the platform is low, the behavior intervals are long, and there are few purchase interaction records. As a result, the products recommended by the recommendation system have a recommendation effect that is significantly weaker than the head recommendation effect.

[0072] Therefore, e-commerce platforms need to improve personalized, precise, and applicable e-product recommendation services for long-tail users with low platform behavior frequency, long behavior intervals, and little purchase interaction record data, thereby increasing the platform stickiness for this type of users and reducing user churn rate.

[0073] The present application provides a method for generating user interest features, a method for clustering user interest features, a method for estimating user behavior objects, a method for training a user behavior object prediction model, and related devices, electronic devices, and storage media, aiming to provide a method for providing accurate product recommendations for long-tail users; naturally, the method of the present application can also be applied to other occasions in the Internet field where object recommendations need to be implemented. The present application does not limit the specific application field. It is only for the convenience of description. The following takes the e-commerce platform recommending suitable products to users as a typical application scenario, and explains the above-mentioned method provided by the present application around this scenario.

[0074] To facilitate understanding of the method embodiment of the present application, its application scenario is introduced. Please refer to Figure 1, which is a schematic diagram of the application scenario of the method embodiment of the present application. It can be applied to a cloud service end. This application scenario is a schematic illustration and does not serve as a specific description of its application scenario.

[0075] As shown in Figure 1, in this application scenario, there are provided a client 101 and a cloud server 102. In this embodiment, the client 101 and the cloud server 102 are directly connected via network communication.

[0076] Client 101, also referred to as the user end, corresponds to cloud server 102 and is controlled and managed by cloud server 102. In the implementation of this embodiment, client 101 can be understood as a client device running a specific application, such as a mobile phone or tablet. It can also be more specifically understood as an application (APP) used to provide local services to users. Specifically, client 101 can include a variety of intelligent terminal devices connected to the Internet, which are numerous. Client 101 and cloud server 102 are typically directly connected via the Internet, thereby providing the computing power and storage capacity of cloud server 102 to the user of client 101. In this embodiment, client 101 is primarily used to obtain user behavior data, such as click behavior data, browsing behavior data, and purchase behavior data. Because user behavior data can be processed and converted into user behavior samples, and user behavior samples are associated with behavior objects, it is possible to analyze the user's interest characteristics by collecting the user's user behavior data, such as obtaining products that the user is interested in and products that the user is not interested in.

[0077] The cloud server 102 includes a cloud server, which can be implemented by virtualizing many physical servers and can expand computing and storage capabilities at any time as needed. The cloud server 102 deploys models related to user interest feature mining. In the embodiment of the present application, these models include the UBP model, the UBC model, and the HIM model. Specifically for the e-commerce platform application scenario, the cloud server 102 also deploys an e-commerce platform. The e-commerce platform provided by the cloud server 102 can receive user behavior data transmitted by the client 101 and perform sample processing on the cloud server 102 to form user behavior samples. These user behavior samples are provided to the above-mentioned model for analysis and processing, and the interest features of the corresponding client 101 users can be obtained.

[0078] In the embodiment of the present application, the cloud service end 102 can be individually set with a UBP model, a UBC model, and a HIM model; the cloud service end 102 can also set the above three models in a comprehensive manner. The setting method can be adjusted according to specific actual needs, and this embodiment does not specifically limit this. The specific physical devices used by the server of the cloud service end, as well as the communication relationship between these devices, can be the responsibility of the manufacturer providing cloud computing services. The cloud service end 102 includes a highly scalable server cluster for virtualization of physical servers. For users, there is no need to care about the specific hardware implementation of the cloud service end 102.

[0079] Based on this, the following describes the method embodiment provided by this application. Please refer to Figure 2, which is a flow chart of the first embodiment of this application. This method embodiment is applied to the basic interest feature expression acquisition model (UBP model). The method is specifically described below in conjunction with Figure 2.

[0080] S201, obtaining a behavior object feature vector corresponding to a target time window for characterizing a user's behavior object; the behavior object feature vector for characterizing the user's behavior object includes a positive feedback object feature vector and a negative feedback object feature vector.

[0081] This step is used to obtain the user's behavior object feature vector.

[0082] The time window is a requirement of the basic interest feature expression acquisition model (UBP model). It is a continuous time segment dynamically preset according to the time parameter to divide the user behavior sequence sample. Since user behavior is based on the timeline, the order of behavior occurrence presents a multi-peak expression. Therefore, the time window set by the UBP model is a dynamic window. The setting of the above-mentioned time window can be adjusted accordingly according to the specific behavior sequence sample of the user. Generally, one or more time windows will be set according to the specific behavior sequence sample of the user. Therefore, the target time window is one of the many time windows divided above. The acquisition of the behavior object feature vector corresponding to the target time window for characterizing the user's behavior object is to obtain the positive feedback object feature vector and the negative feedback object feature vector of a certain user in a time window; further, the behavior object feature vector corresponding to the target time window for characterizing the user's behavior object is a comprehensive feature vector expression of all behavior objects in the time window. In other words, the behavior object feature vector in the above-mentioned target time window is the vector corresponding to multiple behavior objects in the target time window, which is uniformly processed into one behavior object feature vector.

[0083] For example, multiple shopping behaviors of user A's shopping account in the past year concentrated in December. Then, based on the sequence samples of user A's shopping behavior throughout the year, the time window of the current year can be dynamically set for user A, and multiple time windows are set with high density in December, while other months are set with dispersed and sparse settings. Furthermore, the time window corresponding to December 1 is selected as the target time window, and the positive feedback object feature vector and negative feedback object feature vector for characterizing the behavior object of user A under the target time window of December 1 are obtained. The above-mentioned positive feedback object feature vector is a comprehensive vector expression of all positive feedback objects under the target time window of December 1, and the negative feedback object feature vector is a comprehensive vector expression of all negative feedback objects under the target time window of December 1. Schematically, the vectors corresponding to all positive feedback products under the target time window of December 1 form a positive feedback object feature vector.

[0084] The user is the user of the application. In this embodiment, the user can be a user of a variety of applications, such as shopping, office, gaming, and entertainment. The relevant actions performed by the user in the application are called user behavior. Examples include purchase behavior, click behavior, non-click behavior, browsing behavior, adding to shopping cart behavior, and favorite behavior. For ease of understanding, the embodiments of this application use users in shopping applications as examples.

[0085] The user's behavior object refers to the object that can be displayed to the user on the client side of the application. For example, all the products displayed on the display screen of a shopping app of user A. It should be understood that when user A uses the shopping app, all the displayed products can correspond to user behaviors. For example, user A sees product X but is not interested in it, and then ignores it. The above-mentioned ignoring behavior can also be regarded as user behavior of user A in this embodiment. Then, the behavior object of the ignoring behavior is product X.

[0086] Furthermore, the user behavior can be divided into positive feedback behavior and negative feedback behavior. The positive feedback behavior refers to the behavior that the user performs towards the behavior object of interest; the negative feedback behavior refers to the behavior that the user performs towards the behavior object of no interest. That is, any behavior that can reflect the user's interest in the behavior object can be regarded as positive feedback behavior; any behavior that can reflect the user's lack of interest in the behavior object can be regarded as negative feedback behavior. For example, the ignoring behavior mentioned above is a negative feedback behavior corresponding to user A; the click-to-cart behavior can be regarded as the user's positive feedback behavior. Due to the large number of user behavior types, the examples here are illustrative, and those skilled in the art can easily understand the difference between the above-mentioned positive feedback behavior and negative feedback behavior.

[0087] There is an association between user behavior and behavior objects. In other words, user behavior corresponds to a behavior object, and by obtaining the user behavior, the behavior object corresponding to the behavior can be obtained. In this embodiment, this specifically refers to obtaining the user's behavior object and user behavior based on the user behavior sample. The behavior object is determined to be a positive feedback object based on the positive feedback behavior and the association; and the behavior object is determined to be a negative feedback object based on the negative feedback behavior and the association.

[0088] In this embodiment, based on the positive feedback object, the positive feedback object feature vector is obtained; based on the negative feedback object, the negative feedback object feature vector is obtained. The positive feedback object feature vector and the negative feedback object feature vector are both abstract feature expressions in high-dimensional space, and the behavior object feature vector can be used to characterize the abstract features corresponding to the behavior object. In this embodiment, the server on the service side can perform sample processing based on the user behavior data transmitted by the client to obtain user behavior samples. Based on the analysis of the user behavior samples, the positive feedback object feature vector and the negative feedback object feature vector corresponding to the above user behavior data are obtained. Furthermore, based on the above user behavior samples, the positive feedback behavior feature vector and the negative feedback behavior feature vector can also be analyzed. In specific implementation, referring to the diagram in Figure 10, the above user behavior samples can be represented by the symbol "x". After analysis and processing by the server on the service side, the behavior object feature vector "χ" corresponding to the user behavior sample "x" can be obtained. i " and behavioral feature vector Since the behavioral feature vector This is not a research focus in the embodiments of this application, so it will not be elaborated on in detail. The above-mentioned behavior object feature vector includes a positive feedback object feature vector and a negative feedback object feature vector.

[0089] For ease of understanding, in this embodiment, the positive feedback object feature vector and the negative feedback object feature vector are illustrated by way of example. For example, as shown in FIG4 , the abstract identifier of the positive feedback object feature vector is The abstract identifier of the negative feedback object feature vector is For a more detailed description of the process of obtaining the behavior object and behavior object feature vector corresponding to the user behavior, and obtaining the positive feedback object feature vector and negative feedback object feature vector, please refer to the subsequent description.

[0090] The step of obtaining a behavior object feature vector corresponding to a target time window and used to characterize a user's behavior object includes:

[0091] Step S201-1, obtaining user behavior sequence samples;

[0092] Step S201-2, grouping the user behavior sequence samples into corresponding preset time windows to obtain user behavior group samples corresponding to each time window;

[0093] Step S201-3, corresponding to each of the user behavior group samples, obtaining a behavior object corresponding to the user behavior;

[0094] Step S201 - 4 : within the selected target time window, generating the behavior object feature vector corresponding to the target time window according to the behavior object corresponding to the user behavior.

[0095] Next, each of the above steps will be described in detail one by one.

[0096] In step S201 - 1 , a user behavior sample sequence is obtained.

[0097] This step collects samples according to the time sequence of user behaviors, and records user behavior samples in time sequence to form serialized user behavior samples, namely user behavior sequence samples.

[0098] Here, we explain user behavior sequence samples. In the embodiments of this application, the data generated by user behavior is processed on the server side as user behavior samples, such as click behavior data samples (user behavior sample 1), browsing behavior data samples (user behavior sample 2), ... purchase behavior data samples (user behavior sample x), etc. Due to the diversity and temporal sequence of user behavior, user behavior samples can be formed into user behavior sequence samples corresponding to the user according to the timeline based on the chronological order of the user behavior.

[0099] For ease of understanding, please refer to FIG3, which is a schematic diagram of a user A behavior sequence sample corresponding to this embodiment. As shown in FIG3, multiple user behavior samples of user A are represented in the figure as Since user A has multiple user behavior samples, user behavior sample 1 to user behavior sample x, a behavior sequence sample of user A arranged according to the timeline in FIG3 can be formed.

[0100] In step S201 - 2 , the user behavior sequence samples are grouped according to preset time windows to obtain user behavior group samples corresponding to each time window.

[0101] In this step, the user behavior sequence samples collected in the above step S201-1 are grouped according to preset time windows to obtain user behavior group samples corresponding to each time window.

[0102] The grouping process is specifically a process based on the time window. As mentioned above, the time window is a window dynamically preset by the UBP model according to the time parameter for dividing user behavior sequence samples.

[0103] Continuing with the diagram in Figure 3, user A's behavior sequence samples can be grouped according to multiple preset time windows to obtain N user behavior group samples corresponding to the N time windows. For example, based on multiple user behavior samples generated by user A's shopping account over the past year, a user behavior sequence sample for user A is generated, and based on the N preset time windows, N groups of user behavior group samples corresponding to the N time windows are obtained.

[0104] Each specific time window corresponds to user behavior data collected within the time range of the time window, and the user behavior sequence sample collected in each specific time window is regarded as a user behavior group sample.

[0105] The step S201 - 3 is to obtain a behavior object corresponding to each user behavior group sample.

[0106] This step is used to obtain the behavior object corresponding to the user behavior in the user behavior group sample based on the user behavior group sample. That is, for each user behavior group sample, the behavior object corresponding to the user behavior sample can be obtained based on the user behavior sample in the group sample and the behavior object involved in the user behavior. For example, a user behavior sample records a user clicking on product A, then in the user behavior sample, the behavior object corresponding to the user behavior clicking on product A is the behavior object corresponding to the user behavior sample.

[0107] For example, if user A's shopping account recently logged in and a purchase of product X occurred, user behavior sample 1 is generated based on the data generated by this purchase behavior. Based on the user behavior sample and the associated relationship of this purchase behavior, the behavior object is determined to be product X. A user behavior sample includes at least the user ID, user behavior (click, browse, purchase, etc.), behavior object (such as the specific product clicked, browsed, purchased, etc.), and the time of occurrence. The time of occurrence can be used to categorize the sample into a specific time window. Within the selected target time window, the behavior object corresponding to each user behavior sample can be determined based on the content contained in each of the user behavior samples.

[0108] The step S201 - 4 is to generate the behavior object feature vector corresponding to the target time window according to the behavior object corresponding to the user behavior within the selected target time window.

[0109] This step performs mathematical processing, specifically vectorization processing, on the behavior object corresponding to the user behavior determined in the previous step for subsequent use.

[0110] The user behavior corresponds to the behavior object, and the behavior object further corresponds one-to-one to the user's behavior object feature vector.

[0111] The above-mentioned user's behavior object feature vector is an abstract feature expression of the high-dimensional space of the behavior object, which can be used to characterize the abstract features corresponding to the behavior object. If the behavior object is not expressed as an abstract behavior object feature vector, it is impossible to perform various mathematical processing in subsequent models to obtain the patterns therein. For example, the behavior object is "XX mobile phone model: A1". This text-expressed behavior object cannot abstract the hidden features therein with other behavior objects. However, after it is expressed using a vector, through the subsequent training process, the regularity of the association between them can be found.

[0112] The behavior object feature vector includes a positive feedback object feature vector and a negative feedback object feature vector. For some behavior objects, if they are clicked or browsed, they are regarded as positive feedback objects; for some behavior objects, although they are displayed to the user, but the user does not click, purchase, or browse, but directly turns the page, they are regarded as negative feedback objects; after the above-mentioned behavior objects are vectorized and expressed as feature vectors, they become positive feedback object feature vectors and negative feedback object feature vectors. For ease of understanding, please refer to the diagram of Figure 4. Figure 4 is a diagram of the UBP model architecture corresponding to the first embodiment of the present application. As shown in Figure 4, the user's positive feedback object feature vector is represented in the figure by Indicates; the user's negative feedback object feature vector is represented by express.

[0113] In the above steps, after obtaining the behavior objects corresponding to the user behaviors in each of the user behavior group samples, a behavior object feature vector corresponding to a specific selected target time window is generated based on the behavior objects corresponding to the user behaviors. In other words, a corresponding feature vector is generated based on the behavior objects corresponding to the user behaviors to achieve a mathematical representation of the behavior objects.

[0114] The step of generating the behavior object feature vector corresponding to the target time window according to the behavior object corresponding to the user behavior within the selected target time window includes:

[0115] Step S201-4-1, filtering the user's behavior objects in the selected target time window according to preset filtering conditions to obtain sample behavior objects;

[0116] Step S201-4-2: Generate the behavior object feature vector corresponding to the target time window based on the sample behavior object.

[0117] Among them, the preset screening conditions include one or more of the following factors: the frequency of the user's behavior on the behavior object reaches a preset value, the duration of the user's behavior on the behavior object reaches a preset value, and the type of the user's behavior on the behavior object meets the preset type.

[0118] In this embodiment, the filtering of user behavior objects within the selected target time window includes filtering positive feedback objects and negative feedback objects based on the type of the behavior object. For example, the behavior objects corresponding to user behavior group sample 1 for user A in time window 1 (target time window) are obtained. Based on the behavior object type, the positive feedback objects corresponding to positive feedback behaviors are filtered from high to low based on behavior frequency, and a preset behavior frequency threshold is set to obtain the top n positive feedback sample behavior objects corresponding to the preset value. Alternatively, the negative feedback objects corresponding to negative feedback behaviors are filtered based on the length of time that the object has been added to the favorites and added to the shopping cart. Negative feedback objects that meet the preset time threshold are determined as negative feedback sample behavior objects. Through this preliminary screening of behavior objects, the obtained sample behavior objects have a high degree of credibility in the process of generating user interest features. Examples of the preset categories include purchase, add to cart, click to browse, favorite, or share with other users. The preset filtering conditions and preset categories can be combined and adjusted according to the needs of real-world scenarios. This embodiment is provided as an illustrative example and is not intended to be limiting.

[0119] During specific implementation, according to the type of sample behavior objects obtained, positive feedback sample behavior objects and negative feedback sample behavior objects, each generates a behavior object feature vector corresponding to the user behavior group sample corresponding to the target time window for characterizing the user's behavior object, and accordingly, obtains a positive feedback object feature vector and a negative feedback object feature vector.

[0120] In an embodiment of the present application, the above-mentioned process of generating a behavioral object feature vector corresponding to the target time window is implemented in the Input layer module. The Input layer module extracts features from the behavioral object used to characterize the user, and constructs a behavioral object feature vector for characterizing the user's behavior object. The relevant data of the behavior object of the user behavior object pre-stored in the UBP model is input into the input layer for processing, cleaning, dimensionality reduction and other data processing to generate the behavior object feature vector. In this embodiment, the above-mentioned Input layer module can be set in the UBP model, or it can be pre-selected and placed in front of the UBP model. The setting method can be adjusted according to specific actual needs, and this embodiment does not make specific restrictions on this. There are many specific implementation methods for the above-mentioned input layer. For example, it can be understood as an encoding process, which is implemented by the encoding module; or it can be understood as an embedding layer.

[0121] It's important to understand that the aforementioned process of filtering behavior objects corresponding to user behaviors applies to behavior objects within each target time window. In other words, the user's behavior objects within multiple time windows are processed separately, with the current time window in which the behavior object is being processed being called the target time window. Sample behavior objects corresponding to each time window are obtained through filtering. Furthermore, the aforementioned filtering process filters the user's positive and negative feedback objects within each time window separately.

[0122] S202 : performing integration calculation on the negative feedback object characteristic vector to obtain a negative feedback object integrated vector corresponding to the negative feedback object characteristic vector.

[0123] This step is used to obtain the negative feedback object integrated vector corresponding to the negative feedback object feature vector.

[0124] The integration calculation is to integrate the feature vectors of each negative feedback object obtained in step S201 into a unified vector. Various methods can be used, typically, such as maximum pooling or average pooling. This step can be implemented by the pooling layer in Figure 4.

[0125] Pooling involves mimicking the human visual system to reduce the dimensionality of data, representing the original data with higher-level features. This aims to reduce information redundancy and increase computational speed. Pooling can be performed in various ways, including maximum pooling and average pooling.

[0126] The negative feedback object integrated vector is a concentrated expression of the user negative feedback object feature vectors, that is, each negative feedback object feature vector is pooled to form a vector that can reflect their common properties.

[0127] 4, the abstract identifier of the negative feedback object integrated vector is shown in FIG4 as

[0128] The above-mentioned integrated calculation of negative feedback object feature vectors is completed in the Pooling Layer module of the UBP model. Performing maximum pooling or average pooling operations can obtain the negative feedback object integration vector representing the user

[0129] S203: Calculate the correlation between the negative feedback object integrated vector and each of the positive feedback object feature vectors.

[0130] This step is used to obtain the correlation of the feature vectors of each positive feedback object.

[0131] In specific implementations, the aforementioned correlation is calculated using an attention mechanism based on Euclidean distance to obtain the correlation between the negative feedback object integration vector and each of the positive feedback object feature vectors. For example, the attention mechanism is used to perform a dot product operation on the negative feedback object integration vector and the positive feedback object feature vector, and the dot product result is normalized to obtain the correlation.

[0132] The following explains the Euclidean distance and Similarity attention (attention mechanism). Euclidean distance, also known as Euclidean distance, measures the straight-line distance between two points in Euclidean space. In deep learning, Euclidean distance can be used to measure the degree of difference or correlation between two feature vectors. Similarity attention, an attention mechanism originating from cognitive science, is a mechanism in machine learning that often requires models to receive and process large amounts of data. However, at a specific moment, only a small amount of data is considered important. This data is then given more weight during processing, and the computational mechanism based on this weighting is often referred to as the attention mechanism. Introducing the attention mechanism can reduce the amount of information processed and the required computing resources.

[0133] It's important to understand that since correlation reflects the degree of similarity or proximity between the feature vectors of the positive feedback object and the integrated vector of the negative feedback object, the calculated correlation parameter can reflect the degree of trustworthiness of each positive feedback object feature vector, referred to as confidence. Confidence is the probability that the true value will fall within the allowable error range, centered around the estimated value. Confidence reflects the reliability of the data. Therefore, the correlation parameter obtained above can be used to generate the user's primary interest feature expression in subsequent steps.

[0134] S204 : Generate a primary-order interest feature expression of the user corresponding to the target time window according to the correlation, the positive feedback object feature vector, and the negative feedback object integrated vector.

[0135] This step is used to obtain the user's primary interest feature expression corresponding to the target time window.

[0136] The primary interest feature expression is a phased expression of the user's interest feature, which can reflect the user's interest feature corresponding to the above-mentioned target time window. For example, the behavior object feature vector corresponding to time window 1 obtained in step S201 for characterizing the user's behavior object is the behavior object feature vector of the behavior object of user A's shopping account within one day (time window 1). Then, the primary interest feature expression generated in the above step S204 is the primary interest expression feature of user A within one day (time window 1). The above-mentioned primary interest feature expression, as a phased feature expression, can partially or completely serve as the user's interest expression feature.

[0137] The step of generating the user's primary interest feature expression corresponding to the target time window based on the correlation, the positive feedback object feature vector, and the negative feedback object integrated vector includes the following steps. Referring to FIG. 5 , FIG. 5 is a flowchart of generating the user's primary interest feature expression, including the following steps:

[0138] S204-1, obtaining a correlation index between each of the positive feedback object feature vectors and the negative feedback object integrated vector.

[0139] The correlation can reflect the degree of similarity or proximity between the positive feedback object feature vector and the negative feedback object integrated vector. The correlation index, as a reflection of the degree of similarity or proximity between the two feature vectors, can be used to characterize the proximity between behavioral objects. In this embodiment, the correlation index can be presented as a number ranging from 0 to 1, in a matrix format, or in a list format. Corresponding to the UBP model in Figure 4, the correlation index is calculated in the Similarity Attention Module (attention mechanism module). By inputting each positive feedback object feature vector and the negative feedback object integrated vector into the attention mechanism module, the correlation index corresponding to each positive feedback object feature vector can be obtained. In specific implementations, the negative feedback object integrated vector serves as a reference vector in the calculation of each positive feedback object feature vector. By calculating the correlation index between each positive feedback object feature vector and the negative feedback object integrated vector, the degree of proximity or similarity between each pair of positive feedback object feature vectors can be indirectly determined. The correlation index can be calculated using the cosine similarity algorithm, the Euclidean distance algorithm, or the Pearson correlation coefficient algorithm. Other similarity algorithms can also be used, and this embodiment does not limit this.

[0140] S204-2, taking the positive feedback object feature vector whose correlation index is lower than a preset correlation threshold as an available positive feedback object feature vector;

[0141] Here, the step of screening positive feedback object feature vectors that meet the threshold conditions is explained. When a user's user behaviors are limited, noisy behavior object feature vectors corresponding to user behavior samples can significantly damage the UBP model. In particular, when the number of positive feedback behaviors is small and the number of positive feedback object feature vectors corresponding to a user's positive feedback behaviors is high, the UBP model's personalized performance is low, leading to recommendation fatigue. Therefore, it is currently necessary to screen positive feedback object feature vectors corresponding to positive feedback behaviors with low noise and high quality based on a correlation threshold as usable positive feedback object feature vectors. The correlation index can be used as a screening criterion, with positive feedback object feature vectors below a preset correlation threshold being considered as usable positive feedback object feature vectors with low noise and high quality. The specific value of the correlation threshold can be adjusted based on actual scenario requirements and is not limited in this embodiment. A correlation index below the threshold indicates that the positive feedback object has a low correlation with the negative feedback object and is more likely to be a true positive feedback object.

[0142] S204-3: Generate a primary interest feature expression of the user according to the correlation, the available positive feedback object feature vector, and the negative feedback object integrated vector.

[0143] In this step, generating the user's primary interest feature expression based on the correlation, the available positive feedback object feature vector, and the negative feedback object integrated vector includes:

[0144] assigning a second weight to each of the available positive feedback object feature vectors in a preset manner according to the correlation between the available positive feedback object feature vectors and the negative feedback object integrated vector; wherein the higher the correlation, the lower the weight;

[0145] The negative feedback object integrated vector and the weighted positive feedback object feature vector are used to generate a primary interest feature expression of the user.

[0146] In the process of assigning second weights to each of the available positive feedback object feature vectors in the preset manner, the second weight parameter is derived from the correlation index. Since the correlation index is generated using the negative feedback object integrated vector as a reference, the index corresponding to each positive feedback object feature vector is generated. Therefore, the second weight parameter corresponding to each positive feedback object feature vector can be obtained based on the correlation index.

[0147] The purpose of assigning a second weight to each of the available positive feedback object feature vectors is to, based on the weight, reduce the proportion of available positive feedback object feature vectors with lower second weight parameter values ​​in the user's primary interest feature expression. In other words, the more significantly a positive feedback object differs from a negative feedback object, the more it is considered to reflect the characteristics of the positive feedback object and the greater its contribution to the generation of subsequent primary interest feature vectors.

[0148] In this embodiment, the generating of the user's primary interest feature expression using the negative feedback object integrated vector and the weighted positive feedback object feature vector includes:

[0149] S204-3-1, inputting the weighted feature vectors of each positive feedback object into the first transformation neural network to obtain the positive feedback object integrated vector;

[0150] S204-3-2, fusing the positive feedback object integrated vector and the negative feedback object integrated vector to obtain the primary interest feature expression of the user.

[0151] Referring to the diagram in Figure 4, in this embodiment, the first transformation neural network (Transformer Layer module) is shown in Figure 4 as the module corresponding to the first transformation neural network. Transformer Layer is a deep learning model based on the self-attention mechanism, which learns context and thus potential meanings by tracking relationships in sequence data. In specific implementation, after deep learning in the Transformer Layer module, the above-mentioned weighted positive feedback object feature vectors can be collectively expressed as the positive feedback object integration vector of the user. By weighting the second weight parameter in each positive feedback object feature vector, and then performing information interaction between the weighted positive feedback object feature vectors in the Transformer Layer module, the positive feedback object integration vector corresponding to the target time window can be obtained. The positive feedback object integration vector is an integrated expression of the positive feedback objects of the user in the target time window. In other words, the above-mentioned positive feedback object integration vector can represent the expression of the user's stage-specific interest characteristics.

[0152] In step S204-3-2, the positive feedback object integrated vector and the negative feedback object integrated vector are fused to obtain the primary interest feature expression of the user.

[0153] During specific implementation, the positive feedback object integration vector and the negative feedback object integration vector are input into the concatenation module (Concat module) of the UBP model for fusion processing to generate the primary interest feature expression of the user. Among them, the fusion of the positive feedback object integration vector and the negative feedback object integration vector adopts a concatenation method in this embodiment. Of course, the above-mentioned fusion processing can also be vector addition or multiple vectors forming a matrix. The primary interest feature expression is the interest feature expression corresponding to the target time window. That is, the above-mentioned primary interest feature expression can be used as the user's stage-by-stage interest feature expression. The primary interest feature expression obtained will also be different depending on the user's target time window. As shown in Figure 4, the user's primary interest feature expression can be expressed as

[0154] The concat module (full name: concatenate) is a common method for integrating multi-channel features. It concatenates two or more feature vectors to generate a new feature vector. The new feature vector has an increased feature dimension, while retaining the original feature vector information.

[0155] Before the step of generating the user's primary interest feature expression corresponding to the target time window based on the correlation, the positive feedback object feature vector, and the negative feedback object integrated vector, the following steps may also be included:

[0156] assigning a first weight to a corresponding positive feedback object feature vector according to a frequency of each positive feedback object serving as a user behavior object;

[0157] During specific implementation, by inputting each of the positive feedback object feature vectors into a frequency-weight module (freq-weighted layer) module for processing, the first weight parameter for each available positive feedback object feature vector can be obtained. Referring to the schematic diagram of the UBP model in FIG4 , each available positive feedback object feature vector is processed by the frequency-weight module to obtain the positive feedback object feature vector weighted by the first weight parameter corresponding to each positive feedback object feature vector. For example, after the freq-weighted layer module obtains the first weight parameter corresponding to each available positive feedback object feature vector based on each available positive feedback object feature vector, each available positive feedback object feature vector is multiplied by the above-mentioned first weight parameter to obtain the available positive feedback object feature vector weighted by the first weight parameter. Referring to the schematic diagram in FIG4 , the abstract identifier of the available positive feedback object feature vector weighted by the above-mentioned first weight parameter is shown in FIG4 as

[0158] At the same time, in the step of generating the user's primary interest feature expression corresponding to the target time window based on the correlation, the positive feedback object feature vector, and the negative feedback object integrated vector, the available positive feedback object feature vector assigned a first weight is used.

[0159] Here, the purpose of assigning the first weight to the positive feedback object feature vector is to process the available positive feedback object feature vector through the first weight parameter obtained by the freq-weighted layer module (frequency-weight module) to obtain a high-quality available positive feedback object feature vector for generating a primary interest feature expression.

[0160] Through the method provided in the first embodiment above, the primary interest feature expression of a specific user corresponding to a specific time window (target time window) can be obtained; this interest feature expression actually reflects the user interest reflected by the user behavior sample obtained in the target time window.

[0161] Since users may exhibit different interest characteristics in different time windows, in the subsequent method for predicting user behavior targets, the interest characteristics of each time window can be integrated to form a high-level interest characteristic expression of the target user. For details, see the relevant description of the third embodiment.

[0162] Second embodiment

[0163] The following describes an embodiment of the method for clustering user interest features provided by this application in conjunction with Figure 6. Please refer to Figure 6, which is a flow chart of the second embodiment of this application. This method embodiment is applied to the interest group feature expression acquisition model (UBC model), and the method includes the following steps:

[0164] Step S601 , obtaining the target user's primary interest feature expression and interest category feature expression; the interest category feature expression is a feature expression of each interest category formed by classifying users according to their interests.

[0165] This step is used to obtain the primary interest feature expression and interest category feature expression. In specific implementation, the high-level interest feature expression of the target user can also be obtained.

[0166] When the target user's user behaviors are few, sparse, and with large intervals between behaviors, user interest feature clustering can be achieved by obtaining high-quality, high-confidence primary interest feature expressions and interest category feature expressions provided by the UBC model.

[0167] In order to facilitate understanding of this embodiment, reference may also be made to FIG7 , which shows an architecture diagram of the interest group feature expression acquisition model (UBC model) used in this method. The UBC model as a whole is a semi-supervised autoencoder, which is correspondingly provided with an encoder (Encoder) and a decoder (Decoder). Different from the common autoencoder, in the specific implementation, the semi-supervised autoencoder also needs to obtain the interest category feature expression and the target user's high-order interest feature expression in order to realize the processing of user interest feature clustering.

[0168] The target user is a user who has a need for interest category clustering. In other words, the target user is a user who obtains the interest category feature expression corresponding to the interest group to which the target user belongs based on the target user's primary interest feature expression.

[0169] The interest category feature expression obtained by the UBC model is described in detail here. The above-mentioned interest category feature expression used for user interest feature clustering is the interest category feature expression obtained by the UBC model through learning and training. That is, the above-mentioned interest category feature expression obtained is the feature expression initialized randomly; usually, in the UBC model training process, the interest category feature expression initialized randomly and the encoder and decoder parameters randomly initialized are continuously adjusted to finally obtain the interest category feature expression, encoder parameters and decoder parameters used for user interest feature clustering. In other words, the interest category feature expression, encoder parameters and decoder parameters initialized randomly need to be continuously learned and adjusted before they can be used in the process of user interest feature clustering. In specific implementation, the interest category feature expression initialized randomly is used as the input of the UBC model. Through the UBC model training, the interest category feature expression initialized randomly is continuously learned and adjusted, and the encoder and decoder parameters of the above-mentioned UBC model are adjusted to finally obtain the interest category feature expression, encoder parameters and decoder parameters applied to user interest feature clustering.

[0170] The target user's primary interest feature expression is a phased expression of the target user's interest feature, which can reflect the user's interest feature corresponding to the target time window. Generating the target user's primary interest feature expression includes:

[0171] Obtaining a behavior object feature vector corresponding to a target time window for characterizing a behavior object of a target user; the behavior object feature vector for characterizing a behavior object of the target user includes a positive feedback object feature vector and a negative feedback object feature vector;

[0172] Performing integration calculation on the negative feedback object characteristic vector to obtain a negative feedback object integrated vector corresponding to the negative feedback object characteristic vector;

[0173] Calculating the correlation between the negative feedback object integrated vector and each of the positive feedback object feature vectors;

[0174] A primary-order interest feature expression of the target user corresponding to the target time window is generated according to the correlation, the positive feedback object feature vector, and the negative feedback object integrated vector.

[0175] For the specific implementation process of generating the target user's primary interest feature expression corresponding to the target time window, please refer to the description of the first embodiment of this application, which will not be repeated here.

[0176] During specific implementation, the source of the interest category feature expression is the interest category feature expression corresponding to the interest group formed by the UBC model based on the multiple user interest data of multiple users and the preset type division and learning adjustment of the multiple user interest data in the UBC model. That is, the UBC model can divide the above-mentioned multiple user interest data into groups through the collected multiple user interest data of multiple users to form interest category feature expressions with initial random settings corresponding to multiple interest categories. By learning and adjusting the interest category feature expression with initial random settings, an interest category feature expression applied to the interest feature clustering of the target user is finally formed. Referring to the diagram in Figure 7, in the UBC model, a total of K groups (K group) of interest categories are formed, including interest category 1, interest category 2, ... interest category K, and the K interest types correspond to the K group interest category feature expressions (K group embedding). In Figure 7, the vector feature expression "λ" corresponding to the K group interest category feature expression on the right is represented, and the interest feature expression code "λ" of the target user is composed of the interest feature expression code "λ". 0 ”…“λ K-1 ", etc. The interest feature expression encoding "λ" matrix is ​​composed of K groups of interest category feature expression vectors, and the dimension of each vector is D. The above K is a predefined parameter and is determined according to needs; the dimension D of the vector feature expression "λ" is also a parameter that needs to be predefined and is determined according to needs, for example, 32, 64, and 128 are selected as the specific values ​​of dimension D. It should be noted that the data dimensions K and D expressed by the corresponding "λ" matrix are the same as the dimensionality reduction feature expression "β" described in the subsequent steps. i ” and the target user’s interest feature expression code “μ i "The data dimensions are kept consistent. For example, the above interest feature expression code "λ" has K groups of interest category feature expression vectors, corresponding to the reduced dimension feature expression "β i The data dimension of " is also K, and the dimension of each interest category feature expression vector of "λ" is D, corresponding to the target user's interest feature expression code "μ i The data dimension of " is also D. The one-to-one correspondence of the feature expression data dimensions here is to facilitate the one-to-one mapping of the vectors according to the feature dimensions of the vectors when the vectors corresponding to each feature expression are mapped and fused.

[0177] To facilitate understanding of the above process of obtaining interest category feature expressions, an example is given. The UBC model analyzes multiple user interest data and obtains three types of user interests: interest category 1 - promotions, interest category 2 - fitness equipment, and interest category 3 - movies. The interest category feature expressions representing the above three categories are expressed as "λ 1 "、Interest category feature expression "λ 2"、Interest category feature expression "λ 3 ", forming the overall interest category feature expression "λ". Furthermore, the above interest category feature expression "λ" is obtained through learning and adjustment of the interest category feature expression with the initial random setting. It should be emphasized that the interest category feature expression with the initial random setting itself can be a completely meaningless way of dividing interest categories. It is only for the purpose of having a starting point. The actual division method of the final interest category feature expression is entirely the result of machine learning through a large number of samples. Of course, it is not ruled out that the initial interest category division as a starting point has a certain meaning, but this meaning does not really limit the final interest category. It only affects the adjustment process of obtaining the final interest category, and the meaning of this artificial setting is not necessarily a positive impact.

[0178] Step S602 : performing mapping and fusion processing on the primary interest feature expression of the target user according to the interest category feature expression to obtain the interest feature expression code of the target user.

[0179] This step is used to obtain the target user's interest feature expression code.

[0180] Corresponding to the diagram in FIG7, the above primary interest characteristics of the target user are expressed in FIG7. The interest category feature expression is marked with "λ". The encoder converts the above By reducing the amount of data, the high-dimensional data is lossily mapped into low-dimensional data, and the encoded dimensionality reduction feature expression "β i ”. Then the dimensionality reduction feature is expressed as “β i ” is mapped and fused with the interest category feature expression “λ” to obtain the target user’s interest feature expression code “μ i The above mapping fusion processing is the fusion between vectors. The specific processing method can adopt the weighted summation operation and other methods provided by the existing technology. For example, the following formula is used for mapping fusion processing, wherein Corresponding to the above-mentioned encoded dimensionality reduction feature expression "β i "In the hth dimension of the i-th time window (the vector length is k), The feature expression "λ" corresponding to the above interest category is in the s-th row of the i-th time window, and the length of the vector is D.

[0181] In specific implementation, the dimensionality reduction feature expression "β i ” is a feature expression containing the third weight parameter. That is, the dimensionality reduction feature expression “β i " can reflect the feature expression weight.

[0182] The above target user’s interest feature expression code “μ i " can represent the interest characteristics of the target user, and the interest characteristics of the target user are expressed in the code "μ i ” is a fusion of interest category feature expression “λ” and dimensionality reduction feature expression “β i "The feature expression of ". In this embodiment, the dimensionality reduction feature expression after the encoder (Encoder) is "β i The data dimension of " corresponds to the data dimension of K groups of interest category feature expressions "λ" in the UBC model. At the same time, the dimension D of each group of interest category feature expressions corresponds to the target user's interest feature expression code "μ i " has one-to-one correspondence with the data dimensions. In other words, the encoded dimensionality reduction feature expression "β i ” is K dimensions, and the corresponding interest category feature expression “λ” has K groups, each with D dimensions, and the target user’s interest feature expression code “μ i ” is D dimensions. In specific implementation, the UBC model can pre-set the number of groups of user interest category feature expressions and the dimensions of each group, that is, the number of interest category groups and the length of each group vector, and then the encoded dimensionality reduction feature expression “β i "The dimension of the dimensionality reduction feature can be determined. i ” is mapped and fused with the interest category feature expression “λ” to obtain the target user’s interest feature expression code “μ i The specific process is: according to the one-to-one correspondence between data dimensions, the dimensionality reduction feature expression "β i "The third weight parameter and vector on each dimension are respectively multiplied and summed with the vector of the interest category feature expression "λ". Schematically, the "β i The third weight parameter of the dimensionality reduction feature expression of the i-th time window "h=1" (data dimension is the first dimension) At the same time, obtain the interest category feature vector of the sth row in the i-th time window in the interest category feature expression "λ" First, the weight parameters of each group above Multiply each group of interest feature vectors Then, add the corresponding dimensions of the interest category feature expression vector after weight multiplication, and finally get the D-dimensional target user interest feature expression code "μ i ”.

[0183] Step S603: Encode the interest characteristic expression of the target user and perform decoding processing to obtain a decoded interest characteristic expression code.

[0184] This step is used to obtain the feature expression of the interest category after decoding.

[0185] It should be noted that this step is different from the usual encoding-decoding process. In the specific implementation, the feature of interest is expressed in the encoding “μ i During the decoding process, the learning direction of decoding is to express the high-level interest characteristics of the target user. In this step, the “μ i The decoding learning direction of "is set as the high-level interest feature expression of the target user In this way, the interest clustering results can be close to the high-order interest feature expression of the target user, and a better clustering effect can be obtained. The decoded interest feature expression code obtained can better reflect the characteristics of the target user as a user belonging to a certain category.

[0186] In specific implementation, the above target user’s interest feature expression code “μ i ” is performed in the decoder. By increasing the amount of data, the low-dimensional target user’s interest feature expression is encoded “μ i ”, mapped into high-dimensional decoded interest feature expression code “θ i ”.

[0187] Step S604, calculate the loss difference between the decoded interest feature expression encoding and the preset target interest feature expression, and adjust the encoding model, decoding model, and interest category feature expression according to the loss difference until the loss difference reaches the predetermined target; and based on the dimensionality reduction feature expression at this time, obtain the interest category expression to which the target user belongs under the target time window.

[0188] This step is used to obtain the interest category expression to which the target user belongs.

[0189] Among them, the target interest feature expression adopts the user's primary interest feature expression, or adopts the high-order interest feature expression; the high-order interest feature expression is obtained after the primary interest feature expression of the target time window is associated with the primary interest feature expressions of other time windows.

[0190] In this embodiment, specifically: the loss function is used to calculate the encoding "μ i "The decoding interest feature expression code obtained after decoding"θ i " and higher-level interest expressions The loss difference between the two values ​​is used to obtain the above loss difference. The loss difference can be used to measure the feature expression information lost due to compression. The loss difference reaches a predetermined target when the loss difference is less than a preset threshold. According to the diagram in FIG7 , when the loss difference is less than the preset threshold, the dimensionality reduction feature expression “βi ”, obtain the interest category expression “C i ”. Specifically, according to the dimensionality reduction feature expression at this time, “β i ”, get the interest category expression “C i The process of " is shown as follows: According to the dimensionality reduction feature expression "β i "The row vector of the interest category feature expression corresponding to the dimension with the largest vector value in the vector; for example, the dimensionality reduction feature expression at this time is "β i The data dimension of " is 3. During the model training process, the dimensionality reduction feature expression "β i In one round of learning and training, the vector values ​​corresponding to the three data dimensions are respectively "the value of the first dimension is 0.1, the value of the second dimension is 0.8, and the value of the third dimension is 0.2". By comparing the values, we can see that the value of the second dimension is the largest. Therefore, the second row of the interest category feature expression "λ" vector is used as the output result of the UBC model, that is, the interest category expression "C" to which the target user belongs in the target time window. i ”. Further, the above-obtained interest category expression “C i "It can also be used as input data in the UBC model training process to participate in subsequent calculations until the training converges.

[0191] What needs to be understood is that the primary interest feature expression and high-level interest feature expression is a feature expression with a progressive relationship. The above high-order interest feature expression Based on the primary interest feature expression Obtained feature expression. High-order interest feature expression It not only includes the interest features of the time window, but also considers the factors of other time windows, that is, for any target time window, the high-order interest feature expression That is, the user interest features reflected in this time window are considered, and the user interest features reflected in other time windows are also considered to a certain extent. By calculating the decoded interest feature expression code and high-order interest feature expression The difference in loss between the two, the obtained interest category expression “C i "It can be closer to the interest characteristics of the target user.

[0192] The high-order interest feature expression is obtained by associating the primary interest feature expression of the target time window with the primary interest feature expression of other time windows; specifically, the primary interest feature expression of each time window can be correlated with the primary interest feature expression of the target time window by the second transformation neural network 905 shown in FIG9 . Processing is performed to form corresponding high-order interest feature expressions for each time window As mentioned above, the high-order interest feature expression reflects the interest features of this time window and also absorbs the interest features of the user in other time windows.

[0193] In the step S604, the preset target interest feature expression adopts high-order interest feature expression to achieve a semi-supervised effect. Of course, it is not ruled out that the primary interest feature expression is directly used, which is an unsupervised effect. Each time window has a UBC module, and different time windows will learn a user interest cluster expression because user interests will change in different time windows.

[0194] Third embodiment

[0195] The following, in conjunction with Figure 8 , describes an embodiment of a method for estimating user behavior objects provided by this application. Please refer to Figure 8 , which is a flowchart of the third embodiment of this application. This method embodiment is applied to a user behavior object prediction model (HIM model); please refer to Figure 9 for the HIM model. As can be seen, the HIM model includes the UBP model and the UBC model.

[0196] Step S801: Obtain user behavior samples divided according to time windows.

[0197] This step is used to obtain user behavior samples, which are the basis for this method to analyze the interest characteristics of a specific user. These user behavior samples are a series of behavior samples of a user arranged in chronological order. Therefore, in general, they can be called user behavior sequence samples.

[0198] The meaning of the user behavior sample shown has been specifically explained in the description of step S201-1 in the first embodiment of the present application in conjunction with Figure 3 and will not be repeated here.

[0199] The meaning of the time window has been introduced in detail in the description of step S201 of the first embodiment and will not be repeated here.

[0200] Step S802: Generate a behavior object feature vector for characterizing the user's behavior object based on the user behavior sample; the behavior object feature vector for characterizing the user's behavior object includes a positive feedback object feature vector and a negative feedback object feature vector.

[0201] For the process of generating a user behavior object feature vector from the user behavior sample, please refer to the description of step S201 in the first embodiment. In this embodiment, only one wording of step S201 in the first embodiment is changed to emphasize that the behavior object feature vector is obtained from the user behavior sample. The specific implementation of the behavior object generating the behavior object feature vector is formed through the processing of input layer 903 in Figure 9, and this formation process has been described in the first embodiment above.

[0202] Step S803: Generate the user's primary interest feature expression corresponding to each time window based on the behavior object feature vector.

[0203] This step is a summary of the first embodiment; it is specifically performed by the UBP model 904 in Figure 9. It should be noted that the generated primary interest feature expression corresponds to each time window, that is, for each time window in which the user behavior sample is obtained, a corresponding primary interest feature expression is generated. Specifically:

[0204] This step can be further divided into the following specific steps:

[0205] Step S803-1, obtaining a behavior object feature vector corresponding to a target time window for characterizing a user's behavior object; the behavior object feature vector for characterizing the user's behavior object includes a positive feedback object feature vector and a negative feedback object feature vector;

[0206] Step S803-2, performing integration calculation on the negative feedback object characteristic vector to obtain a negative feedback object integrated vector corresponding to the negative feedback object characteristic vector;

[0207] Step S803-3, calculating the correlation between the negative feedback object integrated vector and each of the positive feedback object feature vectors;

[0208] Step S803 - 4 : generating a primary interest feature expression of the user corresponding to the target time window according to the correlation, the positive feedback object feature vector, and the negative feedback object integrated vector.

[0209] The above specific process has been described in detail in the first embodiment and will not be repeated here.

[0210] Step S804 : interactively processing the primary interest feature expression of each time window with the primary interest feature expressions of other time windows to obtain a high-order interest feature expression of each time window.

[0211] In this step, corresponding to each time window, a high-order interest feature expression is generated based on the primary interest feature expression. The high-order interest feature expression is obtained by providing the primary interest feature expression to the second transformation neural network 905 for transformation. The high-order interest feature expression not only includes the interest features of the time window, but also takes into account the factors of other time windows. That is, for any target time window, the high-order interest feature expression takes into account the user interest features reflected in the time window, and also considers the user interest features reflected in other time windows to a certain extent. Of course, the interest features of the target time window are still the main focus. The second transformation neural network 905 (Transformer Layer) can be implemented using a Transformer model; the Transformer model is a neural network that learns context and therefore meaning by tracking relationships in sequence data; the Transformer model applies a set of evolving mathematical techniques called attention or self-attention to detect implicit relationships between even distant data elements in the series that influence and depend on each other.

[0212] The process and principle of generating the high-level interest feature expression from the primary interest feature expression are also described in step S604 of the second embodiment, which can be referred to. After interactive processing, the high-order interest feature expression corresponding to time window 1 is obtained Primary interest characteristic expression After interactive processing, the high-order interest feature expression corresponding to the time window L is obtained

[0213] Step S805, using the primary interest feature expression and high-order interest feature expression of each time window, and dividing the users into types according to their interests to form the interest category feature expression of each interest category, obtain the time window interest category expression to which the target user belongs under each corresponding time window.

[0214] This step obtains the primary interest feature expression, high-order interest feature expression, and interest category feature expression through the previous steps, and obtains the time window interest category expression of the user in the corresponding time window.

[0215] This step is basically a summary of the second embodiment, the difference is that this step emphasizes the interest category expression obtained corresponding to the time window, and the obtained result is called the time window interest category expression, so as to distinguish it from the interest category expression of the target user formed by fusing the interest category expressions of each time window in the subsequent steps. In specific implementation, the primary interest feature expression corresponding to time window 1 and high-level interest feature expression The two are input into the UBC model for processing to obtain the interest category expression of the target user corresponding to time window 1, “C 1 Similarly, under different time windows of the target user, the interest category expression “C 1 、C i …C L ”.

[0216] Specifically, this step may include the following specific steps:

[0217] Step S805-1, obtaining the primary interest feature expression and interest category feature expression of the target user corresponding to the target time window;

[0218] Step S805-2, mapping and fusing the target user's primary interest feature expression according to the interest category feature expression to obtain the target user's interest feature expression code;

[0219] Step S805-3, encoding the target user's interest characteristic expression and decoding the encoding to obtain a decoded interest characteristic expression code;

[0220] Step S805-4, calculate the loss difference between the decoded interest feature expression code and the high-order interest feature expression of the user in the target time window, and adjust the encoding model, decoding model, and interest category feature expression according to the loss difference until the loss difference reaches the predetermined target; and based on the dimensionality reduction feature expression at this time, obtain the interest category expression to which the target user belongs in the target time window.

[0221] For the specific process and meaning of the above steps, please refer to the description of the second embodiment of this application.

[0222] Step S806: The target interest category feature expression of the time window to which the target user belongs in each time window and the user's high-order interest feature expression corresponding to each time window are fused to obtain the target interest category feature expression of the target user.

[0223] This step is implemented by the fusion processing module 907; its purpose is to combine the time window target interest category expression "C i ", and the user's high-level interest feature expression corresponding to each time window Obtain the target interest category feature expression that reflects the overall situation of the target user. Through this step, the interest category that fully reflects the situation of the target user can be obtained. Among them, the fusion processing module 907 can adopt a splicing method. In specific implementation, by expressing the high-level interest feature under each different time window And the target interest category feature expression "C1 …C L " are jointly input into the fusion processing module 907 for processing, thereby obtaining the target interest category feature expression of the target user as output by the module 907. The target interest category feature expression can reflect the interest characteristics of the target user. In other words, the target interest category feature expression can serve as the attribute characteristics of the target user's own interests and reflect the user's interests and hobbies.

[0224] Step S807 : estimating the behavior probability of the target user towards the target behavior object using the target interest category feature expression of the target user.

[0225] This step estimates the target user's behavior probability toward the target behavior object based on the target interest category feature expression obtained in the previous step. Because the target user's target interest category feature expression represents the target user's interest category, the target user's behavior can be estimated using the data of other users in that interest category. This fully utilizes all collected data without being limited by the target user's limited historical behavior data.

[0226] The specific implementation method of this step can be implemented using the following steps. It should be noted that the following method is only one of the methods, and there are obviously other implementation methods in the existing technology:

[0227] Step S807-1, performing attention mechanism processing on the target interest category expression of the target user and the target behavior object feature expression to obtain the attention mechanism expression of the target user.

[0228] In this step, the target action object is the action object for which the target user is evaluated for clicks or other valid actions. For example, a product can be displayed based on this evaluation. As shown in Figure 9, the target action object 902 input into the HIM model is specifically a target action object feature expression.

[0229] The attention mechanism expression of the target user is obtained by processing the target interest category expression of the target user and the characteristic expression of the target behavior object, thereby obtaining the association between the target interest category to which the target user belongs and the target behavior object.

[0230] The attention mechanism processing can be implemented through the attention mechanism processing module 908 (Attention) in Figure 9.

[0231] Step S807-2: Provide the attention mechanism expression of the target user and the characteristic expression of the target behavior object to a pre-trained machine learning model to obtain the behavior probability of the target user towards the target behavior object.

[0232] Since the attention mechanism expression of the target user obtained in the previous step has already expressed the association relationship between the target behavior object and the target interest category described by the target user, this step uses the attention mechanism expression of the target user to calculate the probability of an effective operation occurring with the target behavior object. This step can be implemented in a variety of ways, typically using multiple hidden layers (Multiple Hidden Layer); for example, using the multiple hidden layers 909 in Figure 9. Since the output value range of the multiple hidden layers 909 is large, a normalization step can be further added, that is, the normalization processing of the output results is implemented through the normalization module 910 in Figure 9, and the output probability value is normalized to the range of 0 to 1.

[0233] This embodiment incorporates the UBP model used to implement the method of the first embodiment, and the UBC model used to implement the method of the second embodiment. Based on these two models, a user behavior object prediction model (HIM) is implemented. This model extracts user interest characteristics based on historical user data and clusters them into appropriate interest categories. Using these interest categories, the model estimates the probability that a target user will effectively operate on a target behavior object (e.g., a product on an e-commerce platform). By clustering target users into interest categories and then estimating their behavior probabilities based on their interest categories, this method effectively addresses the problem of predicting the behavior of long-tail users with insufficient data.

[0234] Although the first to third embodiments described above are all methods, they can also be described as models. It can be considered that the methods are implemented by corresponding models, and the models are used to implement the corresponding methods. The two are two sides of the same coin, just different ways of expression.

[0235] The basic processing process of the above HIM model and the method provided in the third embodiment can also be referred to Figure 10, which briefly describes the conversion between feature vectors in the HIM model. As shown in Figure 10, the user behavior sample obtained in a certain time window can be represented by the symbol "x". The above models can analyze and process the user behavior sample "x" to obtain the user's behavior object feature vector "χ" that represents the user's behavior object. i ”, and the behavioral feature vector The above behavior object feature vector "χ i"Including positive feedback object feature vector and negative feedback object feature vector, the positive feedback object feature vector and negative feedback object feature vector are processed by the UBP model to obtain the user's primary interest feature expression Among them, the abstract representation in Figure 10 is Among them, the coefficient "ub" can be understood as the weight parameter of the feature expression. Of course, through the user's behavior feature vector It is also possible to obtain primary interest feature expressions Since this process is not the research focus of this embodiment, it will not be elaborated in detail.

[0236] The primary interest feature expression output by the above UBP model The primary interest feature can be input into the second transformation neural network 905 and the UBC model for processing. Input to the second transformation neural network 905 to obtain the high-order interest feature expression output by the neural network Among them, the coefficient "ub" can also be understood as the weight parameter of the feature expression. Input into the UBC model for encoding, mapping and fusion processing to obtain the low-dimensional target user's interest feature expression encoding Then encode the interest feature expression After decoding, the decoded interest feature expression code "ubθ i ", and the high-order interest feature expression output by the second transformation neural network 905 is expressed as As the direction of clustering learning, the robustness of the end-to-end clustering process is enhanced by constructing a self-supervised / semi-supervised learning process, and the decoded interest feature expression code "ubθ i " and high-level interest feature expression The distance between them is minimized and the similarity between them is maximized. Through the above process, the transformation between the feature vectors in the HIM model is achieved.

[0237] The above HIM model, as well as the UBP model and UBC model included therein, all require extensive training to implement. The following fourth embodiment provides a method for training the HIM model.

[0238] Fourth embodiment

[0239] The fourth embodiment of the present application provides a training method for a user behavior object prediction model (that is, a HIM model). Please refer to Figure 11, and combine it with Figures 9 and 10. Figure 11 is a flow chart of the training method for the user behavior object prediction model of the fourth embodiment of the present application. It should be noted that the training method provided in this embodiment also involves the model principle of the HIM model, that is, the user behavior object prediction method, so reference can be made to the specific description of the third embodiment of the present application. Of course, the specific training details of the UBP model and the UBC model involved in the above-mentioned model training process can also be referred to the description of the first and second embodiments of the present application, which will not be repeated in this embodiment. Furthermore, for the training method of the user behavior object prediction model of the HIM model, end-to-end model training can be achieved by comparing the estimated results of the behavior probability of the behavior object with the actual records, and then completing the adjustment of the relevant parameters in the model.

[0240] Step S1101: Obtain user behavior samples divided according to time windows, including behavior objects involved in the user behavior.

[0241] Step S1102 , corresponding to each time window, generates a behavior object feature vector for characterizing the user's behavior object in the input layer of the existing model; the behavior object feature vector for characterizing the user's behavior object includes a positive feedback object feature vector and a negative feedback object feature vector.

[0242] Step S1103 : generating, in a user interest extraction model, primary interest feature expressions of the user corresponding to each time window based on the behavior object feature vector.

[0243] Step S1104 , interactively processing the primary interest feature expression of each time window and the primary interest feature expressions of other time windows through a transformation neural network to obtain a high-order interest feature expression corresponding to each time window.

[0244] Step S1105, using the primary interest feature expression and high-order interest feature expression of each time window, as well as the interest category feature expression of each interest category formed by classifying users according to their interests, obtain the time window interest category expression to which the target user belongs under each time window in the user interest clustering model.

[0245] Step S1106 , fusing the target interest category expressions of the target user in each time window and the user's high-order interest feature expressions corresponding to each time window to obtain the target interest category expressions of the target user.

[0246] Step S1107 : using the target interest category expression of the target user, an estimation result of the behavior probability of the target user for each behavior object involved in the user behavior sample is estimated.

[0247] Step S1108 , comparing the estimation result with the actual record of the user's behavior for the behavior object in the user behavior sample, and adjusting the parameters of each model involved in the user behavior object estimation model according to the comparison result until the predetermined standard is met.

[0248] This embodiment is basically similar to the third embodiment. As a training method, the behavior object involved in the user behavior sample (which can be a positive sample behavior object or a negative sample behavior object) is used as the target behavior object, and the estimation result of the HIM model is compared with the behavior object of the actual user operation, and the loss is calculated based on the comparison result, and the various parameters in the HIM model are adjusted based on the loss result.

[0249] In addition, the UBP model and UBC model can be trained and adjusted accordingly.

[0250] Fifth embodiment

[0251] The fifth embodiment of the present application provides another method for generating user interest features, which is applied to the cloud service end. Please refer to Figure 12 for a schematic diagram, which is a flow chart of the method for generating user interest features in the fifth embodiment of the present application. It should be noted that this embodiment also involves the model primitive of the HIM model, that is, the user behavior object prediction method, so reference can be made to the specific description of the third embodiment of the present application. Of course, reference can also be made to the description of the first and second embodiments of the present application.

[0252] Step S1201, obtain an interest feature request message sent by the client for requesting to obtain user interest features; it should be understood that the interest feature request message is a request message sent by the client of the target user, and the interest feature request message is used to obtain page content associated with the user interest features corresponding to the target user.

[0253] Step S1202: Returning the page content associated with the user interest feature to the client in response to the interest feature request message; wherein the page content associated with the user interest feature is obtained based on the estimated behavior probability of the target user for the target behavior object;

[0254] The estimated probability of the target user's behavior towards the target behavior object is obtained as follows:

[0255] Obtain user behavior samples divided according to time windows; generate a behavior object feature vector for characterizing the user's behavior object based on the user behavior samples; the behavior object feature vector for characterizing the user's behavior object includes a positive feedback object feature vector and a negative feedback object feature vector; generate the user's primary interest feature expression corresponding to each time window based on the behavior object feature vector; interactively process the primary interest feature expression of each time window with the primary interest feature expression of other time windows to obtain a high-order interest feature expression for each time window; use the primary interest feature expression and high-order interest feature expression of each time window to form interest category feature expressions for each interest category formed by classifying users according to their interests, and obtain the time window interest category expression to which the target user belongs under each time window; fuse the time window target interest category expression to which the target user belongs under each time window and the user's high-order interest feature expression corresponding to each time window to obtain the target interest category feature expression of the target user; use the target interest category expression of the target user to estimate the behavior probability of the target user towards the target behavior object.

[0256] In this embodiment, the process of obtaining the page content associated with the user interest feature in response to the interest feature request message on the cloud server can refer to the description of the user behavior object prediction method embodiment in the third embodiment of this application. In specific implementation, based on the estimated probability of the target user's behavior towards the target behavior object, the target behavior object is screened by a conditional threshold. The target behavior object whose estimated behavior probability meets the conditional threshold is selected as the page content associated with the user interest feature for exposure to the target user on the client.

[0257] Sixth embodiment

[0258] The sixth embodiment of the present application provides another method for generating user interest features. For application client, please refer to FIG13 for a schematic diagram. FIG13 is a flow chart of the method for generating user interest features in the sixth embodiment of the present application. It should be noted that this embodiment also involves the model primitive of the HIM model, that is, the user behavior object estimation method, so the specific description of the third embodiment of the present application can be referred to. Of course, the description of the first and second embodiments of the present application can also be referred to.

[0259] Step S1301: Send an interest feature request message to the cloud server to request the user's interest features;

[0260] Step S1302: receiving the page content associated with the user interest feature returned by the cloud server;

[0261] Step S1303: In response to detecting a trigger operation of the target user on a preset page content, displaying page content associated with the user's interest characteristics; or, in response to detecting a trigger operation of the target user on a query object, displaying page content associated with the user's interest characteristics;

[0262] The page content associated with the user interest feature is obtained by estimating the behavior probability of the target user towards the target behavior object by the cloud server according to the method provided in the fifth embodiment.

[0263] It should be understood that in the embodiment of the present application, the preset page content is the content that is exposed and displayed to the target user when the target user initially enters the application. The above preset page content is randomly initialized content and has a low degree of correlation with the user interest characteristics of the target user. Therefore, in the present embodiment, the target user's client can obtain and display the page content associated with the user interest characteristics of the target user based on the target user's triggering operation on the preset page content, such as the click operation or browsing operation on the product on the initialization page. Alternatively, the target user's client can obtain the page content associated with the user interest characteristics of the target user related to the search content based on the target user's triggering operation on the query object, for example, the search operation of intending to search for a certain product in the page content. The page content associated with the user interest characteristics of the target user can refer to the description of the third embodiment and the fifth embodiment of the present application, and will not be repeated here.

[0264] Seventh embodiment

[0265] The above-mentioned first embodiment provides a method for generating user interest features. Correspondingly, the seventh embodiment of the present application also provides a device for generating user interest features. Since the device embodiment is basically similar to the method embodiment, the description is relatively simple. For the details of the relevant technical features and the implementation effects, please refer to the corresponding description of the method embodiment provided above. The following description of the device embodiment is merely illustrative.

[0266] Please refer to FIG14 to understand this embodiment. FIG14 is a block diagram of the device unit for generating user interest features provided in this embodiment. As shown in FIG4 , the user interest feature generating device 1400 provided in this embodiment includes:

[0267] An acquiring unit 1401 is configured to acquire a behavior object feature vector corresponding to a target time window and used to characterize a user's behavior object; the behavior object feature vector used to characterize the user's behavior object includes a positive feedback object feature vector and a negative feedback object feature vector;

[0268] The processing unit 1402 is configured to perform integration calculation on the negative feedback object feature vector to obtain a negative feedback object integrated vector corresponding to the negative feedback object feature vector;

[0269] The calculation unit 1403 is configured to calculate the correlation between the negative feedback object integrated vector and each of the positive feedback object feature vectors;

[0270] The interest feature expression generating unit 1404 is configured to generate a primary interest feature expression of the user corresponding to the target time window according to the correlation, the positive feedback object feature vector, and the negative feedback object integrated vector.

[0271] Eighth embodiment

[0272] The above-mentioned second embodiment provides a method for clustering user interest features. Correspondingly, the eighth embodiment of the present application also provides a device for clustering user interest features. Since the device embodiment is basically similar to the method embodiment, the description is relatively simple. For the details of the relevant technical features and the implementation effects, please refer to the corresponding description of the method embodiment provided above. The following description of the device embodiment is merely illustrative.

[0273] Please refer to FIG15 to understand this embodiment. FIG15 is a block diagram of the device unit for clustering user interest features provided by this embodiment. As shown in FIG15 , the user interest feature clustering device 1500 provided by this embodiment includes:

[0274] Acquisition unit 1501 is used to acquire the target user's primary interest feature expression and interest category feature expression; the interest category feature expression is a feature expression of each interest category formed by classifying users according to their interests;

[0275] The processing unit 1502 is configured to perform mapping and fusion processing on the primary interest feature expression of the target user according to the interest category feature expression to obtain the interest feature expression code of the target user;

[0276] A construction unit 1503 is configured to encode the interest characteristic expression of the target user and perform decoding processing to obtain a decoded interest characteristic expression code;

[0277] The calculation unit 1504 is used to calculate the loss difference between the decoded interest feature expression encoding and the preset target interest feature expression, and adjust the encoding model, decoding model, and interest category feature expression according to the loss difference until the loss difference reaches the predetermined target; and based on the dimensionality reduction feature expression at this time, obtain the interest category expression to which the target user belongs under the target time window.

[0278] Ninth embodiment

[0279] The above-mentioned third embodiment provides a method for estimating user behavior objects. Correspondingly, the ninth embodiment of the present application also provides a user behavior object estimation device. Since the device embodiment is basically similar to the method embodiment, the description is relatively simple. For the details of the relevant technical features and the implementation effects, please refer to the corresponding description of the method embodiment provided above. The following description of the device embodiment is merely illustrative.

[0280] Please refer to FIG. 16 to understand this embodiment. FIG. 16 is a block diagram of a user behavior object prediction device provided in this embodiment. As shown in FIG. 16 , the user behavior object prediction device 1600 provided in this embodiment includes:

[0281] An acquisition unit 1601 is configured to acquire user behavior samples divided according to time windows;

[0282] The first processing unit 1602 is configured to generate a behavior object feature vector for characterizing the user's behavior object based on the user behavior sample; the behavior object feature vector for characterizing the user's behavior object includes a positive feedback object feature vector and a negative feedback object feature vector;

[0283] The second processing unit 1603 is configured to generate a primary interest feature expression of the user corresponding to each time window based on the behavior object feature vector;

[0284] A feature expression acquisition unit 1604 is configured to interactively process the primary interest feature expression of each time window with the primary interest feature expressions of other time windows to obtain a high-order interest feature expression of each time window;

[0285] An interest category classification unit 1605 is configured to use the primary interest feature expressions and high-order interest feature expressions of each time window to classify users into different interest categories according to their interests, and obtain the time window interest category expression to which the target user belongs in each time window.

[0286] A fusion processing unit 1606 is configured to fuse the target interest category expression of the target user in each time window and the user's high-order interest feature expression corresponding to each time window to obtain the target interest category feature expression of the target user;

[0287] The behavior probability acquisition unit 1607 is configured to estimate the behavior probability of the target user towards the target behavior object using the target interest category expression of the target user.

[0288] Tenth embodiment

[0289] The above-mentioned fourth embodiment provides a training method for a user behavior object prediction model. Correspondingly, the tenth embodiment of the present application also provides a user behavior object prediction model training device. Since the device embodiment is basically similar to the method embodiment, the description is relatively simple. For the details of the relevant technical features and the implementation effects, please refer to the corresponding description of the method embodiment provided above. The following description of the device embodiment is merely illustrative.

[0290] Please refer to FIG17 to understand this embodiment. FIG17 is a block diagram of a user behavior object prediction model training device provided in this embodiment. As shown in FIG17 , the user behavior object prediction model training device 1700 provided in this embodiment includes:

[0291] An acquisition unit 1701 is configured to acquire user behavior samples divided according to time windows, including behavior objects involved in the user behavior;

[0292] The first processing unit 1702 is configured to generate, at the input layer of the existing model, a behavior object feature vector for representing the user's behavior object corresponding to each time window; the behavior object feature vector for representing the user's behavior object includes a positive feedback object feature vector and a negative feedback object feature vector;

[0293] A feature expression generating unit 1703 is configured to generate, in a user interest extraction model, a primary interest feature expression of the user corresponding to each time window based on the behavior object feature vector;

[0294] An interactive processing unit 1704 is configured to interactively process the primary interest feature expression of each time window with the primary interest feature expressions of other time windows through a transformation neural network to obtain a high-order interest feature expression corresponding to each time window;

[0295] Category expression acquisition unit 1705 is configured to use the primary interest feature expression and high-order interest feature expression of each time window, as well as the interest category feature expression of each interest category formed by categorizing users according to their interests, to obtain the time window interest category expression to which the target user belongs in each time window in the user interest clustering model;

[0296] A fusion processing unit 1706 is configured to fuse the target interest category expressions of the target user in each time window and the user's high-order interest feature expressions corresponding to each time window to obtain the target interest category expressions of the target user;

[0297] A behavior probability estimation unit 1707 is configured to estimate the target user's behavior probability for each behavior object involved in the user behavior sample using the target user's target interest category expression;

[0298] The parameter adjustment unit 1708 is used to compare the estimation result with the actual record of the user's behavior towards the behavior object in the user behavior sample, and adjust the parameters of each model involved in the user behavior object estimation model according to the comparison result until the predetermined standard is reached.

[0299] Eleventh embodiment

[0300] The above-mentioned fifth embodiment provides a method for generating user interest features. Correspondingly, the eleventh embodiment of the present application also provides a device for generating user interest features, which is applied to the cloud service end. Since the device embodiment is basically similar to the method embodiment, the description is relatively simple. For the details of the relevant technical features and the implementation effects, please refer to the corresponding description of the method embodiment provided above. The following description of the device embodiment is merely illustrative.

[0301] Please refer to FIG18 to understand this embodiment. FIG18 is a block diagram of the device for generating user interest features provided in this embodiment. As shown in FIG18 , the device for generating user interest features 1800 provided in this embodiment includes:

[0302] The acquiring unit 1804 is configured to acquire an interest feature request message sent by the client to request obtaining a user interest feature;

[0303] The transmission unit 1802 is configured to return the page content associated with the user interest feature to the client in response to the interest feature request message; wherein the page content associated with the user interest feature is obtained based on the estimated behavior probability of the target user for the target behavior object;

[0304] The estimated probability of the target user's behavior towards the target behavior object is obtained as follows:

[0305] Obtain user behavior samples divided according to time windows; generate a behavior object feature vector for characterizing the user's behavior object based on the user behavior samples; the behavior object feature vector for characterizing the user's behavior object includes a positive feedback object feature vector and a negative feedback object feature vector; generate the user's primary interest feature expression corresponding to each time window based on the behavior object feature vector; interactively process the primary interest feature expression of each time window with the primary interest feature expression of other time windows to obtain a high-order interest feature expression for each time window; use the primary interest feature expression and high-order interest feature expression of each time window to form interest category feature expressions for each interest category formed by classifying users according to their interests, and obtain the time window interest category expression to which the target user belongs under each time window; fuse the time window target interest category expression to which the target user belongs under each time window and the user's high-order interest feature expression corresponding to each time window to obtain the target interest category feature expression of the target user; use the target interest category expression of the target user to estimate the behavior probability of the target user towards the target behavior object.

[0306] Twelfth embodiment

[0307] The above-mentioned sixth embodiment provides a method for generating user interest features. Correspondingly, the twelfth embodiment of the present application also provides a device for generating user interest features, which is applied to the client. Since the device embodiment is basically similar to the method embodiment, the description is relatively simple. For the details of the relevant technical features and the implementation effects, please refer to the corresponding description of the method embodiment provided above. The following description of the device embodiment is merely illustrative.

[0308] Please refer to FIG19 to understand this embodiment. FIG19 is a block diagram of the device unit for generating user interest features provided in this embodiment. As shown in FIG19 , the user interest feature generating device 1900 provided in this embodiment includes:

[0309] The sending unit 1901 is configured to send an interest feature request message to the cloud server to request obtaining the user interest feature;

[0310] The receiving unit 1902 is configured to receive the page content associated with the user interest feature returned by the cloud server;

[0311] The first processing unit 1903 is configured to display page content associated with a user interest characteristic in response to detecting a trigger operation of the target user on a preset page content; or to display page content associated with the user interest characteristic in response to detecting a trigger operation of the target user on a query object;

[0312] The page content associated with the user interest feature is obtained by estimating the behavior probability of the target user towards the target behavior object by the cloud server according to the device provided in the eleventh embodiment.

[0313] Thirteenth embodiment

[0314] In the above embodiments, a variety of methods and corresponding devices are provided. In addition, the present application also provides electronic device embodiments corresponding to the above method embodiments and device embodiments. Since the electronic device embodiments are basically similar to the method embodiments, the description is relatively simple. For details of the relevant technical features and the effects achieved, please refer to the corresponding description of the above method embodiments. The following description of the electronic device embodiment is merely illustrative. The electronic device embodiment is as follows:

[0315] Please refer to Figure 20 to understand the above-mentioned electronic device. Figure 20 is a schematic diagram of the electronic device. The electronic device provided in this embodiment includes: a processor 2001 and a memory 2002, a communication bus 2003 and a communication interface 2004. The processor 2001 is used to execute computer software instructions; the computer software designed according to the methods provided in the above-mentioned embodiments is stored in the memory 2002. When the instructions of the memory 2002 are loaded into the processor 2001, these instructions are executed to implement the steps of the above-mentioned method embodiments. The communication bus 2003 is used to connect the processor 2001 and the memory 2002 mounted thereon. The communication interface 2004 is used to provide a connection interface for the processor 2001 and the memory 2002.

[0316] Fourteenth embodiment

[0317] In the above embodiments, various methods and corresponding devices are provided. In addition, embodiments of the present application also provide computer-readable storage media for implementing the above methods. The computer-readable storage medium embodiments provided in this application are described in a relatively simple manner. For relevant details, please refer to the corresponding descriptions of the above method embodiments. The embodiments described below are merely illustrative.

[0318] The computer-readable storage medium provided in this embodiment stores computer instructions, which, when executed by a processor, implement the steps shown in the above method embodiment, which will not be repeated here.

[0319] In a typical configuration, a computing device includes one or more processors (CPUs), input / output interfaces, network interfaces, and memory.

[0320] Memory may include non-permanent storage in a computer-readable medium, random access memory (RAM) and / or non-volatile memory in the form of read-only memory (ROM) or flash RAM. Memory is an example of a computer-readable medium.

[0321] 1. Computer-readable media includes permanent and non-permanent, removable and non-removable media that can be implemented by any method or technology to store information. Information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic cassettes, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transmission media that can be used to store information that can be accessed by a computing device. As defined herein, computer-readable media does not include non-transitory media such as modulated data signals and carrier waves.

[0322] 2. Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware. Furthermore, the present application may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0323] Although the present invention is disclosed above in terms of preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art may make possible changes and modifications without departing from the spirit and scope of the present invention. Therefore, the scope of protection of the present invention shall be based on the scope defined by the claims of the present invention.

[0324] It should be noted that the embodiments of this application may involve the use of user data. In actual applications, user-specific personal data can be used in the solutions described herein, subject to the requirements of the applicable laws and regulations of the country in which the application is located (for example, with the explicit consent of the user, effective notification to the user, etc.), and within the scope permitted by applicable laws and regulations. It should be noted that the embodiments of this application may involve the use of user data. In actual applications, user-specific personal data can be used in the solutions described herein, subject to the requirements of the applicable laws and regulations of the country in which the application is located (for example, with the explicit consent of the user, effective notification to the user, etc.), and within the scope permitted by applicable laws and regulations.

Claims

1. A method for generating user interest features, characterized in that: include: Acquire a behavior object feature vector corresponding to the target time window for characterizing the user's behavior object; the behavior object feature vector for characterizing the user's behavior object includes a positive feedback object feature vector and a negative feedback object feature vector; Performing integration calculation on the negative feedback object feature vector to obtain a negative feedback object integrated vector corresponding to the negative feedback object feature vector; Calculating the correlation between the negative feedback object integrated vector and each of the positive feedback object feature vectors; A primary-order interest feature expression of the user corresponding to the target time window is generated according to the correlation, the positive feedback object feature vector, and the negative feedback object integrated vector.

2. The method for generating user interest features according to claim 1, characterized in that: The step of obtaining a behavior object feature vector corresponding to the target time window and used to characterize the user's behavior object includes: Obtain user behavior sequence samples; The user behavior sequence samples are grouped and processed corresponding to the preset time windows to obtain user behavior group samples corresponding to each time window; Corresponding to each of the user behavior group samples, obtaining a behavior object corresponding to the user behavior therein; Within the selected target time window, a behavior object feature vector corresponding to the target time window is generated according to the behavior object corresponding to the user behavior.

3. The method for generating user interest features according to claim 2, characterized in that: The step of generating, within the selected target time window, the behavior object feature vector corresponding to the target time window according to the behavior object corresponding to the user behavior, comprises: According to the preset screening conditions, the behavior objects of the users in the selected target time window are screened to obtain sample behavior objects; According to the sample behavior object, a behavior object feature vector corresponding to the target time window is generated.

4. The method for generating user interest features according to claim 3, characterized in that: The preset screening condition includes one or more of the following factors: the frequency of the user's behavior on the behavior object reaches a preset value, the duration of the user's behavior on the behavior object reaches a preset value, and the type of the user's behavior on the behavior object meets the preset type.

5. The method for generating user interest features according to claim 1, characterized in that: The step of generating the primary interest feature expression of the user corresponding to the target time window according to the correlation, the positive feedback object feature vector, and the negative feedback object integrated vector includes: Obtaining a correlation index between each of the positive feedback object feature vectors and the negative feedback object integrated vector; The positive feedback object feature vector whose correlation index is lower than a preset correlation threshold is used as an available positive feedback object feature vector; A primary interest feature expression of the user is generated according to the correlation, the available positive feedback object feature vector and the negative feedback object integrated vector.

6. The method for generating user interest features according to claim 5, characterized in that: The step of generating the primary interest feature expression of the user according to the correlation, the available positive feedback object feature vector and the negative feedback object integrated vector comprises: According to the correlation between the available positive feedback object feature vector and the negative feedback object integrated vector, a second weight is assigned to each of the available positive feedback object feature vectors in a preset manner; wherein the higher the correlation, the lower the weight; The negative feedback object integrated vector and the weighted positive feedback object feature vector are used to generate a primary interest feature expression of the user.

7. The method for generating user interest features according to claim 6, characterized in that: The step of using the negative feedback object integrated vector and the weighted positive feedback object feature vector to generate the user's primary interest feature expression includes: Inputting each weighted positive feedback object feature vector into a first transformation neural network to obtain a positive feedback object integrated vector; The positive feedback object integrated vector and the negative feedback object integrated vector are fused to obtain the primary interest feature expression of the user.

8. The method for generating user interest features according to claim 1, characterized in that: Before the step of generating the primary interest feature expression of the user corresponding to the target time window according to the correlation, the positive feedback object feature vector, and the negative feedback object integrated vector, the following steps are included: According to the frequency of each of the positive feedback objects being used as a user behavior object, assigning a first weight to the corresponding positive feedback object feature vector; In the step of generating the user's primary interest feature expression corresponding to the target time window based on the correlation, the positive feedback object feature vector, and the negative feedback object integrated vector, the positive feedback object feature vector assigned a first weight is used.

9. According to the method for generating user interest features according to claim 1, the fusion of the positive feedback object integrated vector and the negative feedback object integrated vector is performed in a splicing manner.

10. A method for clustering user interest features, characterized in that: The method comprises: Obtaining the target user's primary interest feature expression and interest category feature expression; the interest category feature expression is a feature expression of each interest category formed by classifying users according to their interests; According to the interest category feature expression, mapping and fusing the primary interest feature expression of the target user is performed to obtain the interest feature expression code of the target user; The interest characteristic expression code of the target user is encoded and decoded to obtain a decoded interest characteristic expression code; Calculate the loss difference between the decoded interest feature expression encoding and the preset target interest feature expression, and adjust the encoding model, decoding model, and interest category feature expression according to the loss difference until the loss difference reaches the predetermined target; and obtain the interest category expression to which the target user belongs under the target time window according to the reduced dimensionality feature expression at this time.

11. The method for clustering user interest features according to claim 10, characterized in that: The target interest feature expression adopts the primary interest feature expression of the user, or adopts the high-order interest feature expression; the high-order interest feature expression is obtained by associating the primary interest feature expression of the target time window with the primary interest feature expressions of other time windows.

12. A method for estimating user behavior objects, characterized in that: include: Obtain user behavior samples divided by time windows; Generate a behavior object feature vector for characterizing the user's behavior object according to the user behavior sample; the behavior object feature vector for characterizing the user's behavior object includes a positive feedback object feature vector and a negative feedback object feature vector; Generating, according to the behavior object feature vector, a primary interest feature expression of the user corresponding to each time window; Interactively process the primary interest feature expression of each time window with the primary interest feature expression of other time windows to obtain a high-order interest feature expression of each time window; Using the primary interest feature expressions and high-order interest feature expressions of each time window, and classifying users into different types according to their interests, interest category feature expressions of each interest category are formed to obtain the time window interest category expressions to which the target user belongs in each time window; The target interest category expression of the time window to which the target user belongs in each time window and the high-order interest feature expression of the user corresponding to each time window are fused to obtain the target interest category feature expression of the target user; The target interest category expression of the target user is used to estimate the behavior probability of the target user towards the target behavior object.

13. The user behavior object prediction method according to claim 12, characterized in that: The step of generating the primary interest feature expression of the user corresponding to each time window according to the behavior object feature vector, wherein the process of generating the primary interest feature expression corresponding to the target time window includes: Acquire a behavior object feature vector corresponding to the target time window for characterizing the user's behavior object; the behavior object feature vector for characterizing the user's behavior object includes a positive feedback object feature vector and a negative feedback object feature vector; Performing integration calculation on the negative feedback object feature vector to obtain a negative feedback object integrated vector corresponding to the negative feedback object feature vector; Calculating the correlation between the negative feedback object integrated vector and each of the positive feedback object feature vectors; A primary-order interest feature expression of the user corresponding to the target time window is generated according to the correlation, the positive feedback object feature vector, and the negative feedback object integrated vector.

14. The method for estimating user behavior objects according to claim 12, characterized in that: The method of using the primary interest feature expression and the high-order interest feature expression of each time window to classify users into different types according to their interests, and obtaining the interest category expression of the time window to which the target user belongs under each time window, includes: Obtaining the target user's primary interest feature expression and interest category feature expression in the target time window; According to the interest category feature expression, mapping and fusing the primary interest feature expression of the target user is performed to obtain the interest feature expression code of the target user; The interest characteristic expression code of the target user is encoded and decoded to obtain a decoded interest characteristic expression code; Calculate the loss difference between the decoded interest feature expression encoding and the high-order interest feature expression of the user in the target time window, and adjust the encoding model, decoding model, and interest category feature expression according to the loss difference until the loss difference reaches a predetermined target; and obtain the interest category expression to which the target user belongs in the target time window according to the reduced dimensionality feature expression at this time.

15. The method for estimating user behavior objects according to claim 12, characterized in that: The estimating the behavior probability of the target user towards the target behavior object by using the target interest category expression of the target user includes: Performing attention mechanism processing on the target interest category expression of the target user and the target behavior object feature expression to obtain the attention mechanism expression of the target user; The attention mechanism expression of the target user and the characteristic expression of the target behavior object are provided to a pre-trained machine learning model to obtain the behavior probability of the target user towards the target behavior object.

16. A method for training a user behavior object prediction model, characterized in that: include: Obtaining user behavior samples divided according to time windows, including behavior objects involved in the user behaviors; Corresponding to each time window, a behavior object feature vector for characterizing the user's behavior object is generated at the input layer of the existing model; the behavior object feature vector for characterizing the user's behavior object includes a positive feedback object feature vector and a negative feedback object feature vector. Feedback object feature vector; According to the behavior object feature vector, generating the primary interest feature expression of the user corresponding to each time window in the user interest extraction model; Interactively processing the primary interest feature expression of each time window with the primary interest feature expression of other time windows through a transformation neural network to obtain a high-order interest feature expression corresponding to each time window; Using the primary interest feature expression and high-order interest feature expression of each time window, and the interest category feature expression of each interest category formed by classifying users according to their interests, in the user interest clustering model, the time window interest category expression to which the target user belongs corresponding to each time window is obtained; The target interest category expression of the time window to which the target user belongs in each time window and the high-order interest feature expression of the user corresponding to each time window are fused to obtain the target interest category expression of the target user; Using the target interest category expression of the target user to estimate the estimated result of the target user's behavior probability for each behavior object involved in the user behavior sample; The estimation result is compared with the actual record of the user's behavior for the behavior object in the user behavior sample, and according to the comparison result, the parameters of each model involved in the user behavior object estimation model are adjusted until a predetermined standard is reached.

17. A method for generating user interest features, applied to a cloud service end, the method comprising: Obtaining an interest feature request message sent by a client for requesting to obtain a user interest feature; In response to the interest feature request message, returning the page content associated with the user interest feature to the client; wherein the page content associated with the user interest feature is obtained based on the estimated behavior probability of the target user to the target behavior object; The estimated behavior probability of the target user to the target behavior object is obtained as follows: Obtain user behavior samples divided according to time windows; generate a behavior object feature vector for characterizing the user's behavior object based on the user behavior samples; the behavior object feature vector for characterizing the user's behavior object includes a positive feedback object feature vector and a negative feedback object feature vector; generate the user's primary interest feature expression corresponding to each time window based on the behavior object feature vector; interactively process the primary interest feature expression of each time window with the primary interest feature expression of other time windows to obtain a high-order interest feature expression of each time window; use the primary interest feature expression and high-order interest feature expression of each time window to form interest category feature expressions of each interest category by classifying users according to user interests, and obtain the time window interest category expression to which the target user belongs under each time window; fuse the time window target interest category expression to which the target user belongs under each time window and the user's high-order interest feature expression corresponding to each time window to obtain the target interest category feature expression of the target user; use the target interest category expression of the target user to estimate the target The probability of a user's behavior towards the target behavior object.

18. A method for generating user interest features, applied to a client, the method comprising: Sending an interest feature request message to the cloud server to request obtaining the user's interest features; Receiving the page content associated with the user interest feature returned by the cloud service end; In response to detecting a trigger operation of a target user on a preset page content, displaying page content associated with the user's interest characteristics; or, in response to detecting a trigger operation of the target user on a query object, displaying page content associated with the user's interest characteristics; Among them, the page content associated with the user interest feature is obtained by the cloud server end according to the method for generating user interest features provided in claim 17, estimating the behavior probability of the target user towards the target behavior object.

19. An electronic device, characterized in that: comprising a processor and a memory; wherein, The memory is used to store one or more computer instructions, wherein the one or more computer instructions are executed by the processor to implement the method according to any one of claims 1-18.

20. A computer-readable storage medium having one or more computer instructions stored thereon, characterized in that: The instructions are executed by a processor to implement the method according to any one of claims 1 to 18.

Citation Information

Patent Citations

  • Content pushing method and device

    CN108681919A

  • Method and device for determining user label, equipment and medium

    CN110516175A

  • Content processing method and device, storage medium and electronic equipment

    CN114036396A

  • Building user profiles by relevance feedback

    US20190171725A1

Cited By

  • E-commerce data updating processing method and system based on state difference real-time feedback

    CN120994672A

  • E-commerce dynamic marketing resource intelligent allocation method and system based on multi-modal user behavior characteristics

    CN122243617A

  • E-commerce dynamic marketing resource intelligent allocation method and system based on user behavior characteristics

    CN122243617B

  • User interest modeling method and device and electronic equipment

    CN122470820A

  • User interest modeling method, apparatus, and electronic device

    CN122470820B