A method, device and apparatus for recommending

Abstract

The recommendation method provided by the present application includes: obtaining user data of a target user and recommendation data of a target object of interest to be recommended to the target user, taking the target user and the target object of interest as calculation objects respectively, obtaining a sub-behavior set for characterizing all interactive behaviors between the calculation object and each interactive object, determining the comprehensive preference degree of the calculation object for each interest dimension in a pre-set interest matrix, for each sub-behavior, performing feature extraction on the sub-behavior based on the comprehensive preference degree of the calculation object for each interest dimension, obtaining a feature vector of the calculation object in each interest dimension in the sub-behavior, and determining a matrix composed of the feature vectors of the calculation object in each interest dimension in each sub-behavior as the interest preference matrix corresponding to the calculation object, and determining whether to recommend the target object of interest to the target user based on the interest preference matrix corresponding to the target user and the interest preference matrix corresponding to the target object of interest.

Description

Technical Field

[0001] The present application relates to the field of recommendation technology, and in particular to a recommendation method, device and equipment. Background Art

[0002] In recent years, with the rapid development of Internet technology, recommendation technology has gradually emerged. In recommendation technology, how to mine and characterize users' diverse interest preferences has become the focus of current research.

[0003] The current mainstream technology is to represent users' interest preferences through low-dimensional embedding vectors. For example, a two-tower model is used to extract the interest preference matrix. However, the two towers of the existing two-tower model are independent of each other, and the feature interaction ability is weak. It cannot effectively mine and characterize diverse interest preferences, resulting in poor recommendation effects and poor user satisfaction. Summary of the Invention

[0004] In view of this, the present application provides a recommendation method, apparatus and device to effectively characterize diverse interests, improve recommendation effects and enhance user satisfaction.

[0005] Specifically, this application is implemented through the following technical solutions:

[0006] A first aspect of the present application provides a recommendation method, the method comprising:

[0007] Acquire user data of a target user and recommendation data of a target object of interest to be recommended for the target user;

[0008] Taking the target user and the target object of interest as computing objects, respectively, obtaining a sub-behavior set representing all interaction behaviors between the computing object and each interactive object; wherein each interactive object is each interactive object in the interactive object set obtained based on user data of the computing object or historical interaction data in recommendation data;

[0009] Determining the comprehensive preference of the computing object for each interest dimension in a preset interest matrix;

[0010] For each sub-behavior, feature extraction is performed on the sub-behavior based on the comprehensive preference degree of the computing object for each interest dimension, and the feature vector of the computing object in each interest dimension in the sub-behavior is obtained, and the matrix composed of the feature vectors of the computing object in each interest dimension in each sub-behavior is determined as the interest preference matrix corresponding to the computing object;

[0011] Based on the interest preference matrix corresponding to the target user and the interest preference matrix corresponding to the target object of interest, it is determined whether to recommend the target object of interest to the target user.

[0012] A second aspect of the present application provides a recommendation device, which includes an acquisition module, a processing module, a calculation module, and a judgment module; wherein:

[0013] The acquisition module is used to acquire user data of a target user and recommendation data of a target object of interest to be recommended to the target user;

[0014] The acquisition module is further configured to use the target user and the target object of interest as computational objects, respectively, and acquire a sub-behavior set representing all interaction behaviors between the computational object and each interactive object; wherein each interactive object is an interactive object in the interactive object set acquired based on user data of the computational object or historical interaction data in recommendation data;

[0015] The processing module is used to determine the comprehensive preference degree of the computing object for each interest dimension in the preset interest matrix;

[0016] The calculation module is used to extract features of each sub-behavior based on the comprehensive preference degree of the calculation object for each interest dimension, obtain the feature vector of the calculation object in each interest dimension in the sub-behavior, and determine the matrix composed of the feature vectors of the calculation object in each interest dimension in each sub-behavior as the interest preference matrix corresponding to the calculation object;

[0017] The judgment module is configured to determine whether to recommend the target object of interest to the target user based on the interest preference matrix corresponding to the target user and the interest preference matrix corresponding to the target object of interest.

[0018] The third aspect of the present application provides a recommendation device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the steps of any one of the methods provided in the first aspect of the present application are implemented.

[0019] The fourth aspect of the present application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of any one of the methods provided in the first aspect of the present application.

[0020] The recommendation method, apparatus, and device provided herein obtain user data of a target user and recommendation data of target objects of interest to be recommended for the target user. Next, using the target user and target objects of interest as computational objects, a set of sub-behaviors is obtained, representing all interactions between the computational object and each object. Each object is a member of the set of objects obtained based on the computational object's user data or historical interaction data in the recommendation data. The method then determines the computational object's preference for each interest dimension in a pre-set interest matrix. Feature extraction is then performed on each sub-behavior based on the computational object's overall preference for each interest dimension, obtaining a feature vector for the computational object in each sub-behavior. A matrix of the feature vectors for each sub-behavior in each interest dimension is then determined as the computational object's corresponding interest preference matrix. Finally, based on the target user's interest matrix and the target object of interest's interest preference matrix, a determination is made as to whether to recommend the target object of interest to the target user. In this way, by constructing sub-behaviors, interest preferences are described from multiple interest dimensions based on fine-grained sub-behaviors, resulting in a more comprehensive characterization of the target user. In addition, through the globally shared interest matrix, the high-level representations of target users and target objects of interest can be mapped to the same interest space and aligned on the interest dimension. This can improve the feature interaction capability, effectively mine and characterize diverse interest preferences, improve recommendation effects, and thus improve user satisfaction. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 This is a flowchart of a recommended method according to an embodiment of the present application;

[0022] Figure 2 This is a schematic diagram illustrating an implementation principle of calculating a comprehensive preference degree according to an exemplary embodiment of the present application;

[0023] Figure 3 A flowchart of a recommended method shown in another embodiment of the present application;

[0024] Figure 4 A flowchart of a recommended method shown in another embodiment of the present application;

[0025] Figure 5 This is a flowchart of a recommended method shown in yet another embodiment of the present application;

[0026] Figure 6 This is a schematic structural diagram of a double-tower model according to an exemplary embodiment of the present application;

[0027] Figure 7 A hardware structure diagram of the recommended device where the recommended device provided in this application is located;

[0028] Figure 8 This is a structural diagram of the first embodiment of the recommended device provided in this application. DETAILED DESCRIPTION

[0029] Exemplary embodiments will be described in detail herein, with examples illustrated in the accompanying drawings. In the following description, when referring to the drawings, identical numerals in different figures represent identical or similar elements, unless otherwise indicated. The embodiments described in the following exemplary embodiments are not intended to represent all embodiments consistent with the present application. Rather, they are merely examples of apparatus and methods consistent with certain aspects of the present application, as detailed in the appended claims.

[0030] The terms used in this application are for the purpose of describing specific embodiments only and are not intended to limit this application. As used in this application and the appended claims, the singular forms "a," "an," "the," and "the" are intended to include the plural forms, unless the context clearly indicates otherwise. It should also be understood that the term "and / or" as used herein refers to and encompasses any and all possible combinations of one or more of the associated listed items.

[0031] It should be understood that although the terms first, second, third, etc. may be used in this application to describe various information, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from each other. For example, without departing from the scope of this application, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Depending on the context, the word "if" as used herein may be interpreted as "at the time of" or "when" or "in response to determining".

[0032] This application provides a recommendation method, apparatus, and device to effectively characterize diverse interests, improve recommendation effects, and enhance user satisfaction.

[0033] The recommendation method, apparatus, and device provided herein obtain user data of a target user and recommendation data of a target object of interest to be recommended to the user. Next, using the target user and the target object of interest as computational objects, a set of sub-behaviors is obtained, representing all interactions between the computational object and each object of interest. Each object of interest is a member of the set of objects obtained based on the computational object's user data or historical interaction data in the recommendation data. The method then determines the computational object's preference for each interest dimension in a pre-set interest matrix. Feature extraction is then performed on each sub-behavior based on the computational object's overall preference for each interest dimension, obtaining a feature vector for the computational object in each sub-behavior. A matrix of the feature vectors for each sub-behavior in each interest dimension is then determined as the interest preference matrix corresponding to the computational object. Finally, based on the interest matrix corresponding to the target user and the interest preference matrix corresponding to the target object of interest, a determination is made as to whether to recommend the target object of interest to the target user. In this way, by constructing sub-behaviors, interest preferences are described from multiple interest dimensions based on fine-grained sub-behaviors, resulting in a more comprehensive characterization of the target user. In addition, through the globally shared interest matrix, the high-level representations of target users and target objects of interest can be mapped to the same interest space and aligned on the interest dimension. This can improve the feature interaction capability, effectively mine and characterize diverse interest preferences, improve recommendation effects, and thus improve user satisfaction.

[0034] Specific embodiments are given below to introduce the technical solutions of the present application in detail.

[0035] Figure 1 This is a flowchart of a recommended method shown in one embodiment of the present application. Figure 1 The recommended method provided in this embodiment may include:

[0036] S101: Obtain user data of a target user and recommendation data of a target object of interest to be recommended to the target user.

[0037] Specifically, the specific data types of user data and recommendation data are determined based on actual needs. In this embodiment, the specific data types included in user data and recommendation data are not limited. For example, in one embodiment, user data may include static attribute information of the user and historical interaction data between the user and an interactive object; recommendation data may include static attribute information of a target object of interest and historical interaction data between the target object of interest and the user.

[0038] It should be noted that the target interest to be recommended for the target user (for ease of distinction, the target user will be referred to as User 1) can be a commodity, an article, music, a social object, etc. In this embodiment, the target interest to be recommended for the target user is not limited. The following description takes the target interest to be recommended for the target user as a commodity as an example. In this case, the user data includes the user's static attribute information and the historical interaction data of the target user's interaction with the commodity (wherein the historical interaction data of the target user's interaction with the commodity includes but is not limited to the interaction data generated by the target user's purchase of commodity 1, the target user's favorite behavior of commodity 2, the target user's evaluation of commodity 3, and other interaction behaviors). Furthermore, in this case, the recommendation data of the target interest includes the static attribute data of the target interest (for example, the target interest is commodity 5) and the interaction data of the target interest and the user (the interaction data includes but is not limited to the interaction data generated by the interaction behaviors such as the purchase of the interest by User 2, the favorite behavior of User 3, and the evaluation of the interest by User 4).

[0039] Furthermore, user data and recommendation data may be obtained in different ways, which are not limited in this embodiment.

[0040] S102: Taking the target user and the target object of interest as computing objects, respectively, obtaining a sub-behavior set representing all interaction behaviors between the computing object and each interactive object; wherein each interactive object is each interactive object in the interactive object set obtained based on user data of the computing object or historical interaction data in recommendation data.

[0041] As described above, historical interaction data refers to data generated by interactions between a computational object and an interactive object. When the computational object is a target user, the historical interaction data refers to data generated by interactions between the target user and the interactive object (e.g., a product) (e.g., purchase, favorite, review, etc.). When the computational object is a target object of interest (e.g., a product), the historical interaction data refers to data generated by interactions between the target object of interest (product) and the interactive object (user) (e.g., purchase, favorite, review, etc.).

[0042] It should be noted that a piece of historical interaction data refers to data generated by an interaction between a computing object and at least one interactive object.

[0043] In this step, for the target user, based on the historical interaction data in the user data of the target user, first obtain the interactive object set 1 consisting of all interactive objects that interact with the target user, and then for each interactive object in the interactive object set 1, obtain the sub-behavior of the target user's interaction with the interactive object, and finally obtain the sub-behavior set 1 corresponding to the target user.

[0044] For the target object of interest, based on the historical interaction data in the recommendation data of the target object of interest, first obtain the interactive object set 2 consisting of all interactive objects that interact with the target object of interest, and then for each interactive object in the interactive object set 2, obtain the sub-behavior of the target object of interest interacting with the interactive object, and finally obtain the sub-behavior set 2 corresponding to the target object of interest.

[0045] For example, in one embodiment, the target interest object to be recommended to the target user (referred to as user 1 for easy distinction) may be social object 1. At this time, the historical interaction data in the user data of the target user obtained includes: user 1 likes user 2 and user 5, user 1 adds user 3 as a friend, and user 1 adds user 4 to the blacklist; accordingly, at this time, the constructed sub-behavior set 1 of the target user includes four sub-behaviors, which are: sub-behavior 11 of user 1 liking user 2, sub-behavior 12 of user 1 liking user 5, sub-behavior 13 of user 1 adding user 3 as a friend, and sub-behavior 14 of user 1 adding user 4 to the blacklist (it should be noted that 11, 12, 13, and 14 are identification information of each sub-behavior in the sub-behavior set 1).

[0046] For another example, in this example, the historical interaction data in the recommendation data obtained for social object 1 includes: social object 1 setting user 6 as a favorite user, social object 1 adding user 7 as a friend, and the social object liking user 9. At this time, the constructed sub-behavior set 2 of social object 1 includes three sub-behaviors, namely: sub-behavior 21 of social object 1 setting user 6 as a favorite user, sub-behavior 22 of social object 1 adding user 7 as a friend, and sub-behavior 23 of the social object liking user 9 (it should be noted that 21, 22, and 23 are the identification information of each sub-behavior in sub-behavior set 2).

[0047] It should be noted that, optionally, in one embodiment, each sub-behavior can be represented by a 1*d vector. For a certain computing object, when the sub-behavior set corresponding to the computing object involves N sub-behaviors, the sub-behavior set can be represented by an N*d matrix.

[0048] Furthermore, a specific example will be given below to introduce the specific implementation process of this step in detail, which will not be repeated here.

[0049] S103: Determine the comprehensive preference degree of the computing object for each interest dimension in a preset interest matrix.

[0050] Specifically, the interest matrix includes multiple interest dimensions. For example, if the target object of interest is a commodity, the interest matrix may include interest dimensions such as price, quality, rating, appearance, and brand. It should be noted that the interest matrix is ​​set based on actual needs, and the specific content of the interest matrix is ​​not limited in this embodiment. When setting up the interest matrix, it is best to ensure that the set interest matrix covers the entire interest space.

[0051] For example, in one embodiment, the interest matrix is ​​represented by an M*1 matrix, each row of the matrix represents an interest dimension, the interest matrix contains a total of M interest dimensions, and each interest dimension can be represented by a 1*d vector.

[0052] In specific implementation, the implementation process of this step may include:

[0053] (1) For each sub-behavior, map the sub-behavior to the interest space corresponding to the interest matrix to obtain a preference vector of the computational object in the sub-behavior in the interest space; wherein each element in the preference vector represents the preference of the computational object in the sub-behavior for each interest dimension in the interest matrix.

[0054] For example, combining with the above example, when the calculation object is the target user, at this time, the target user includes 4 sub-behaviors. In this step, for sub-behavior 11, it is mapped to the interest space corresponding to the interest matrix to obtain the preference vector of the calculation object in the sub-behavior 11 in the interest space. Figure 2 This is a schematic diagram showing an implementation principle of calculating the comprehensive preference level according to an exemplary embodiment of the present application. Figure 2 Specifically, during the mapping, the sub-behavior 11 is first mapped to the interest dimension 0 in the interest matrix, and the preference of the computation object in the sub-behavior 11 for the interest dimension 0 in the interest matrix is ​​obtained. Furthermore, the above process is repeated for each interest dimension in the interest matrix to obtain the preference vector of the computation object in the sub-behavior 11 in the interest space corresponding to the interest matrix. Furthermore, the above process is repeated for other sub-behaviors to obtain the preference vector of the computation object in the interest space for each sub-behavior.

[0055] It should be noted that, combined with Figure 2As in the example above, each sub-behavior can be represented by a 1*d vector. When the sub-behavior set corresponding to the computational object involves N sub-behaviors, the sub-behavior set can be represented by an N*d matrix, and the interest matrix can be represented by an M*1 matrix. Accordingly, during the specific mapping, the correlation matrix between the matrix corresponding to the sub-behavior set and the interest matrix is ​​calculated. Each column of the correlation matrix is ​​the preference vector of the computational object in the interest space for a sub-behavior. Combined with the example above, the correlation matrix is ​​an M*N matrix (each element in the matrix is ​​represented by a 1*d vector).

[0056] Further, combined Figure 2 , in specific implementation, the implementation process of this step can be represented by the following formula:

[0057]

[0058] Among them, R i To calculate the object's preference vector in the interest space in the i-th sub-behavior, Q is the interest matrix, K i is the vector corresponding to sub-behavior i. It is understood that when the interest matrix includes M dimensions, Ri contains M values. For the convenience of the following description, these M values ​​are recorded as Ri1, ..., RiM. Among them, Rij represents the computing object's preference for the jth interest dimension in the interest matrix in the i-th sub-behavior.

[0059] (2) For each interest dimension, the preference of the computing object for the interest dimension in each sub-behavior is aggregated to obtain the comprehensive preference degree of the computing object for the interest dimension.

[0060] Combined with the above example, for N sub-behaviors, the matrix composed of the preference vectors of the computing object in the interest space in all sub-behaviors is an M*N matrix. Each row of the matrix represents the preference of the computing object for a certain interest dimension in each sub-behavior. In this step, each row of the matrix is ​​aggregated to obtain the comprehensive preference degree of the computing object for each interest dimension. Combined with the above example, the matrix composed of the comprehensive preference degree of the computing object for all interest dimensions is represented by an M*1 matrix. Each row of the matrix is ​​represented by a 1*d vector. In other words, the matrix composed of the comprehensive preference degree of the computing object for all interest dimensions is an M*d matrix.

[0061] In specific implementation, for each interest dimension, the aggregation process can be represented by the following formula:

[0062] Aj=softmax(R1j, R2j..., RNj)

[0063] Among them, Aj represents the comprehensive preference of the calculation object for the j-th interest dimension, and RNj is the preference of the calculation object for the j-th interest dimension in the interest matrix in the N-th sub-behavior.

[0064] S104. For each sub-behavior, feature extraction is performed on the sub-behavior based on the comprehensive preference degree of the computing object for each interest dimension, and the feature vector of the computing object in each interest dimension in the sub-behavior is obtained, and the matrix composed of the feature vectors of the computing object in each interest dimension in each sub-behavior is determined as the interest preference matrix corresponding to the computing object.

[0065] Referring to the previous description, the matrix of the comprehensive preference degree of the object for all interest dimensions is expressed as A, A=[A1、A2,……,A M ], combined with Figure 2 , the specific implementation process of this step can be expressed as:

[0066] V mat =A*K

[0067] Among them, V mat is the interest preference matrix corresponding to the calculation object, A is the matrix composed of the comprehensive preference degree of the calculation object for all interest dimensions, and K is the matrix corresponding to the sub-behavior set of the calculation object.

[0068] S105: Determine whether to recommend the target object of interest to the target user based on the interest preference matrix corresponding to the target user and the interest preference matrix corresponding to the target object of interest.

[0069] In the above example, assuming User 1 is the target user and Item 5 is the target object of interest, after obtaining the interest preference matrices corresponding to User 1 and Item 5, a determination is made based on these two interest matrices to determine whether to recommend the target object of interest to the target user. The specific implementation process and principles of this step can be found in the description of the relevant art and will not be elaborated here. For example, in one embodiment, the similarity value between the two interest preference matrices can be calculated, and then, if the similarity value exceeds a preset threshold, the target object of interest is recommended to the target user.

[0070] The recommendation method provided in this embodiment obtains user data and recommendation data for a target object of interest to be recommended to the user. Then, using the user and target object of interest as computational objects, a set of sub-behaviors representing all interactions between the computational object and each object of interest is obtained. Each object of interest is a member of the set of sub-behaviors obtained based on the computational object's user data or historical interaction data in the recommendation data. The method then determines the computational object's preference for each interest dimension in a pre-set interest matrix. Feature extraction is then performed on each sub-behavior based on the computational object's overall preference for each interest dimension, obtaining a feature vector for the computational object in each sub-behavior. The matrix of feature vectors for each sub-behavior in each interest dimension is then used as the computational object's interest matrix. Finally, based on the user's interest matrix and the target object of interest matrix, a decision is made as to whether to recommend the target object of interest to the user. This effectively models the user's complex interests and preferences, effectively resolving the inherent feature filtering issues of deep matching models and enabling targeted recommendations.

[0071] Figure 3 This is a flowchart of a recommended method shown in another embodiment of the present application. Figure 3 Based on the above embodiment, the process of obtaining a sub-behavior set for representing all interaction behaviors between the computing object and each interactive object may include:

[0072] S301. Perform feature extraction on attribute information and historical interaction data in user data or recommendation data of the computing object to obtain an object embedding vector of the computing object and object embedding vectors of each interacting object that interacts with the computing object.

[0073] It should be noted that feature extraction can be performed based on a traditional feature extraction method or a feature extraction method based on a neural network. For example, when a traditional feature extraction method is used for feature extraction, the method may be a behavioral feature extraction method, a statistical feature extraction method, etc. For another example, when a feature extraction method based on a neural network is used, the method may be based on a pre-trained feature extraction network for feature extraction. Furthermore, when feature extraction is performed based on a pre-trained feature extraction network, the attribute information and historical interaction data in the user data or recommendation data of the calculation object may be respectively input into the pre-trained feature extraction network, so as to perform feature extraction on the input data through the feature extraction network, and obtain the object embedding vector of the calculation object and the object embedding vector of each interacting object that interacts with the calculation object.

[0074] S302: For each interactive object, construct a sub-behavior based on the object embedding vector and the object embedding vector of the interactive object.

[0075] In a specific implementation, for each interactive object, the interactive behavior between the computing object and the interactive object is represented by a vector consisting of the object embedding vector of the computing object and the object embedding vector of the interactive object.

[0076] Combined with the above example, for example, for sub-behavior 11 (the behavior of user 1 liking user 2), the sub-behavior 11 can be represented by a vector composed of the object embedding vector of user 1 and the object embedding vector of user 2.

[0077] The method provided in this embodiment provides a method for constructing sub-behaviors. By extracting features from the attribute information and historical interaction data in the user data or recommendation data of the calculation object, the object embedding vector of the calculation object and the object embedding vectors of each interactive object that interacts with the calculation object are obtained. Then, for each interactive object, a sub-behavior is constructed based on the object embedding vector and the object embedding vector of the interactive object. Through this sub-behavior construction method, sub-behaviors can be constructed to facilitate subsequent calculations and recommendations.

[0078] Figure 4 This is a flowchart of a recommended method shown in another embodiment of the present application. Figure 4 The method provided in this embodiment, based on the above embodiment, aggregating the preference of the computing object for the interest dimension in each sub-behavior to obtain the comprehensive preference of the computing object for the interest dimension, may include:

[0079] S401. For each interest dimension, select a specified number of preferences from the preferences of the computing object for the interest dimension in each sub-behavior.

[0080] Specifically, the designated number is set according to actual needs. In this embodiment, the designated data is not limited. When this case is implemented through a model, the designated number is determined after the model is trained.

[0081] In a specific implementation, for each interest dimension, a specified number of preference degrees may be selected from the preference degrees of the calculation object for the interest dimension in each sub-behavior in descending order.

[0082] S402: Aggregate the selected preferences.

[0083] The specific implementation process and implementation principle of this step can be found in the description of the previous embodiment and will not be repeated here.

[0084] The method provided in this embodiment determines the comprehensive preference degree of the computing object for the interest dimension. For each interest dimension, a specified number of preference degrees are selected from the preference degrees of the computing object for the interest dimension in various sub-behaviors, and the selected preference degrees are aggregated to obtain the comprehensive preference degree of the computing object for the interest dimension. In this way, it can be ensured that the sub-behaviors are only activated on the most significant specified number of interests, eliminating the influence of insignificant interests, ensuring the sparsity of subsequent interest aggregation, and realizing effective interaction of feature selection and related information in the matching process of target users and target recommendation objects, so that the interest preference matrix can effectively mine and characterize diverse interest preferences.

[0085] Figure 5 This is a flowchart of a recommended method shown in another embodiment of the present application. Figure 5 The method provided in this embodiment may include:

[0086] S501: Obtain user data of a target user and recommendation data of a target object of interest to be recommended to the target user.

[0087] The specific implementation process and implementation principle of this step can be found in the description of the previous embodiment and will not be repeated here.

[0088] S502: Taking the target user and the target object of interest as computing objects, respectively, obtaining a sub-behavior set representing all interaction behaviors between the computing object and each interactive object; wherein each interactive object is each interactive object in the interactive object set obtained based on user data of the computing object or historical interaction data in recommendation data.

[0089] S503. For each sub-behavior, map the sub-behavior to the interest space corresponding to the interest matrix to obtain a preference vector of the computational object in the interest space in the sub-behavior; wherein each element in the preference vector represents the preference of the computational object for each interest dimension in the interest matrix in the sub-behavior.

[0090] S504. For each interest dimension, select a specified number of preferences from the preferences of the computing object for the interest dimension in each sub-behavior, and aggregate the selected preferences to obtain a comprehensive preference degree of the computing object for the interest dimension.

[0091] S505. For each sub-behavior, feature extraction is performed on the sub-behavior based on the comprehensive preference degree of the computing object for each interest dimension, and the feature vector of the computing object in each interest dimension in the sub-behavior is obtained, and the matrix composed of the feature vectors of the computing object in each interest dimension in each sub-behavior is determined as the interest matrix corresponding to the computing object.

[0092] Specifically, in a possible implementation, the process of obtaining the interest preference matrix of the computing object is based on the two towers in the pre-trained two-tower model. For example, Figure 6 This is a schematic diagram of a double tower model according to an exemplary embodiment of the present application. Figure 6 The dual-tower model includes a first tower and a second tower. The first tower is used to obtain the interest preference matrix corresponding to the user based on the input user data, and the second tower in the dual-tower model is used to obtain the interest preference matrix corresponding to the target interest object based on the input recommendation data.

[0093] Further, refer to Figure 6 The dual-tower model is further provided with an interest matrix layer, in which the interest matrix is ​​pre-configured. The interest matrix layer is respectively connected to the first tower and the second tower for providing the interest matrix to the first tower and the second tower.

[0094] It should be noted that, referring to Figure 6 In the specific implementation, in step S502, the user data of the target user is input into the first tower of the dual-tower model, and the recommendation data of the target interest object is input into the second tower of the dual-tower model. Then, the first tower of the dual-tower model outputs the interest preference matrix of the target user, and the second tower of the dual-tower model outputs the interest preference matrix of the target interest object.

[0095] For details, please refer to Figure 6 The first tower and the second tower of the dual-tower model include a feature extraction layer and a sub-behavior construction layer. In step S502, after the user data is input into the first tower and the recommendation data is input into the second tower, the feature extraction layer can extract features from the attribute information and historical interaction data in the input user data or recommendation data respectively, and obtain the object embedding vector of the calculation object (target user or target recommendation object) and the object embedding vector of each interacting object that interacts with the calculation object.

[0096] Furthermore, the sub-behavior construction layer constructs a sub-behavior for each interactive object based on the object embedding vector and the object embedding vector of the interactive object.

[0097] Please continue to refer to Figure 6 , the first tower and the second tower also include a multi-interest attention mechanism layer, and the multi-interest attention mechanism layer is connected to the interest matrix layer and the sub-behavior construction layer respectively; wherein,

[0098] The multi-interest attention mechanism layer is specifically used to determine the preference degree of the computing object for each interest dimension in the interest matrix, and for each sub-behavior, perform feature extraction on the sub-behavior based on the comprehensive preference degree of the computing object for each interest dimension, obtain the feature vector of the computing object in the sub-behavior under each interest dimension, and determine the matrix composed of the feature vectors of each sub-behavior under each interest dimension as the interest preference matrix corresponding to the computing object.

[0099] When implementing it, refer to Figure 6 , a multi-interest attention mechanism layer, specifically used to map each sub-behavior to the interest space corresponding to the interest matrix, and obtain the preference vector of the computing object in the sub-behavior in the interest space; wherein each element in the preference vector represents the preference of the computing object for each interest dimension in the interest matrix in the sub-behavior;

[0100] The multi-interest attention mechanism layer is also specifically used to aggregate the preference of the computing object for each interest dimension in each sub-behavior to obtain the comprehensive preference degree of the computing object for the interest dimension.

[0101] In specific implementation, the multi-interest attention mechanism layer is mainly used to execute steps S503 to S505. The specific implementation process and implementation principle of steps S503 to S505 can be found in the previous description and will not be repeated here.

[0102] It should be noted that, see Figure 6 ,After obtaining the interest preference matrix, the interest preference matrix can be ,vectorized, and then the recommendation is decided based on the vectorized ,interest preference matrix.

[0103] S506: Determine whether to recommend the target object of interest to the target user based on the interest preference matrix corresponding to the target user and the interest preference matrix corresponding to the target object of interest.

[0104] For details, please refer to Figure 6 The dual-tower model also includes a matching calculation layer. This layer is connected to the multi-interest attention mechanism layer in the dual-tower model. Specifically, the matching calculation layer is used to obtain the interest preference matrices output by the first and second towers, perform similarity calculations on the two interest preference matrices, and obtain a similarity value. This similarity value is then used to determine whether to make recommendations.

[0105] It should be noted that when training the twin-tower model, a specified number k needs to be trained. The selection method of the specified number k is particularly important when training the model. When the specified number k is set to a smaller specific value, the difficulty of training the model is greatly increased, and the model is almost unable to converge. Therefore, during specific training, a guided learning training method can be used for training to simulate the learning process from easy to difficult. The specified number k is initially set to a larger value to simplify the training task and quickly make the model reach a certain reliability. As the training progresses, the value of the specified number k is gradually reduced, and the difficulty of training the model is gradually increased until the model reaches a relatively stable convergence state. Through this model training method, both the rapid convergence of the model and the reliability of the model are guaranteed.

[0106] The method provided in this embodiment has at least the following advantages:

[0107] (1) Through the globally shared interest matrix, the high-level representations of target users and target objects of interest can be mapped to the same interest space and aligned in the interest dimension. This can improve the feature interaction capability, effectively mine and characterize diverse interest preferences, improve recommendation effects, and thus improve user satisfaction.

[0108] (2) Sub-behaviors are activated only on the most significant specified number of interests, eliminating the influence of insignificant interests, ensuring the sparsity of subsequent interest aggregation, and realizing effective interaction between feature selection and related information in the matching process between target users and target recommendations, so that the interest preference matrix can effectively mine and characterize diverse interest preferences.

[0109] (3) During the model learning process, only the sub-behavior sets corresponding to the interest dimensions activated on both sides will participate in the final matching calculation and model parameter update. That is, for the same user's high-level preference representation, when different target interests participate in model training, different sub-behaviors in the user's sub-behavior set will participate in this model training process, so that feature screening can be effectively performed.

[0110] Corresponding to the aforementioned embodiment of a recommendation method, the present application also provides an embodiment of a recommendation device.

[0111] The embodiment of a recommendation device of the present application can be applied to recommendation. The device embodiment can be implemented by software, or by hardware or a combination of software and hardware. Taking software implementation as an example, as a device in a logical sense, it is formed by the processor of the recommendation device where it is located reading the corresponding computer program instructions in the non-volatile memory into the memory and running it. From the hardware level, such as Figure 7 The figure shows a hardware structure diagram of the device recommended by this application, except Figure 7In addition to the processor, memory, network interface, and non-volatile memory shown, the recommended device where the apparatus is located in the embodiment may also include other hardware according to the actual function of the recommended apparatus, which will not be described in detail.

[0112] Figure 8 This is a schematic diagram of the structure of the first embodiment of the recommended device provided by this application. Please refer to Figure 8 The device provided in this embodiment may include an acquisition module 810, a processing module 820, a calculation module 830 and a judgment module 840; wherein,

[0113] The acquisition module 810 is configured to acquire user data of a target user and recommendation data of a target object of interest to be recommended to the target user.

[0114] The acquisition module 810 is further configured to use the target user and the target object of interest as computing objects, respectively, to acquire a sub-behavior set representing all interaction behaviors occurring between the computing object and each interactive object; wherein each interactive object is an interactive object in the interactive object set acquired based on the user data of the computing object or historical interaction data in the recommendation data.

[0115] The processing module 820 is used to determine the comprehensive preference degree of the computing object for each interest dimension in a preset interest matrix.

[0116] The calculation module 830 is used to extract features of each sub-behavior based on the comprehensive preference degree of the calculation object for each interest dimension, obtain the feature vector of the calculation object in each interest dimension in the sub-behavior, and determine the matrix composed of the feature vectors of the calculation object in each interest dimension in each sub-behavior as the interest preference matrix corresponding to the calculation object.

[0117] The judgment module 840 is configured to determine whether to recommend the target object of interest to the target user based on the interest preference matrix corresponding to the target user and the interest preference matrix corresponding to the target object of interest.

[0118] The device of this embodiment can be used to perform Figure 1 The steps, specific implementation principles and implementation processes of the illustrated method embodiment can be referred to the description in the previous embodiments and will not be repeated here.

[0119] Optionally, the processing module 820 is specifically configured to map, for each sub-behavior, the sub-behavior to the interest space corresponding to the interest matrix, and obtain a preference vector of the computing object in the sub-behavior in the interest space; wherein each element in the preference vector represents the preference of the computing object in the sub-behavior for each interest dimension in the interest matrix;

[0120] The processing module 820 is further configured to aggregate, for each interest dimension, the preference of the computing object for the interest dimension in each sub-behavior to obtain a comprehensive preference degree of the computing object for the interest dimension.

[0121] Furthermore, the processing module 820 is specifically configured to select, for each interest dimension, a specified number of preferences from the preferences of the calculation object for the interest dimension in each sub-behavior, and aggregate the selected preferences.

[0122] Optionally, the acquisition module 810 is specifically used to perform feature extraction on the attribute information and historical interaction data in the user data or recommendation data of the computing object, respectively, to obtain the object embedding vector of the computing object and the object embedding vector of each interacting object that interacts with the computing object, and for each interacting object, to construct a sub-behavior based on the object embedding vector and the object embedding vector of the interacting object.

[0123] Optionally, the process of obtaining the interest preference matrix of the calculation object is obtained based on two towers in a pre-trained two-tower model; wherein the first tower of the two-tower model is used to obtain the interest preference matrix corresponding to the target user based on the input user data, and the second tower of the two-tower model is used to obtain the interest preference matrix corresponding to the target interest object based on the input recommendation data;

[0124] The dual-tower model is further provided with an interest matrix layer, in which the interest matrix is ​​pre-configured. The interest matrix layer is connected to the first tower and the second tower respectively, and is used to provide the interest matrix to the first tower and the second tower.

[0125] Optionally, the first tower and the second tower further include a feature extraction layer and a sub-behavior construction layer, wherein,

[0126] The feature extraction layer is used to extract features from the attribute information and historical interaction data in the user data or recommendation data of the computing object, respectively, to obtain the object embedding vector of the computing object and the object embedding vectors of each interactive object interacting with the computing object;

[0127] The sub-behavior construction layer is used to construct a sub-behavior for each interactive object based on the object embedding vector and the object embedding vector of the interactive object.

[0128] Optionally, the first tower and the second tower further include a multi-interest attention mechanism layer, which is connected to the interest matrix layer and the sub-behavior construction layer respectively; wherein,

[0129] The multi-interest attention mechanism layer is specifically used to determine the comprehensive preference degree of the computing object for each interest dimension in the interest matrix, and for each sub-behavior, perform feature extraction on the sub-behavior based on the comprehensive preference degree of the computing object for each interest dimension, obtain the feature vector of the computing object in the sub-behavior under each interest dimension, and determine the matrix composed of the feature vectors of each sub-behavior under each interest dimension as the interest preference matrix corresponding to the computing object.

[0130] The implementation process of the functions and effects of each unit in the above-mentioned device is specifically described in the implementation process of the corresponding steps in the above-mentioned method, and will not be repeated here.

[0131] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to the partial description of the method embodiments. The device embodiments described above are merely schematic, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the present application scheme. A person of ordinary skill in the art can understand and implement it without paying any creative work.

[0132] The above description is only a preferred embodiment of the present application and is not intended to limit the present application. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present application shall be included in the scope of protection of the present application.

Claims

1. A recommendation method, characterized in that: The method comprises: Acquire user data of a target user and recommendation data of a target object of interest to be recommended for the target user; Taking the target user and the target object of interest as computing objects, respectively, obtaining a sub-behavior set representing all interaction behaviors between the computing object and each interactive object; wherein each interactive object is each interactive object in the interactive object set obtained based on user data of the computing object or historical interaction data in recommendation data; Determining the comprehensive preference of the computing object for each interest dimension in a preset interest matrix; For each sub-behavior, feature extraction is performed on the sub-behavior based on the comprehensive preference degree of the computing object for each interest dimension, and the feature vector of the computing object in each interest dimension in the sub-behavior is obtained, and the matrix composed of the feature vectors of the computing object in each interest dimension in each sub-behavior is determined as the interest preference matrix corresponding to the computing object; Based on the interest preference matrix corresponding to the target user and the interest preference matrix corresponding to the target object of interest, it is determined whether to recommend the target object of interest to the target user.

2. The method according to claim 1, characterized in that Determining the comprehensive preference of the computing object for each interest dimension in a preset interest matrix includes: For each sub-behavior, mapping the sub-behavior to the interest space corresponding to the interest matrix, and obtaining a preference vector of the computing object in the sub-behavior in the interest space; wherein each element in the preference vector represents the preference of the computing object in the sub-behavior for each interest dimension in the interest matrix; For each interest dimension, the preference of the computing object for the interest dimension in each sub-behavior is aggregated to obtain the comprehensive preference degree of the computing object for the interest dimension.

3. The method according to claim 2, characterized in that The step of aggregating the preference of the computing object for the interest dimension in each sub-behavior to obtain the comprehensive preference of the computing object for the interest dimension includes: For each interest dimension, selecting a specified number of preference degrees from the preference degrees of the computing subject for the interest dimension in each sub-behavior; Aggregate the selected preferences.

4. The method according to claim 1, wherein The obtaining of a sub-behavior set for characterizing all interaction behaviors between the computing object and each interactive object includes: Performing feature extraction on attribute information and historical interaction data in user data or recommendation data of the computation object to obtain an object embedding vector of the computation object and object embedding vectors of each interacting object with the computation object; For each interactive object, a sub-behavior is constructed based on the object embedding vector and the object embedding vector of the interactive object.

5. The method according to claim 4, characterized in that The process of obtaining the interest preference matrix of the calculation object is based on the two towers in a pre-trained two-tower model; wherein the first tower of the two-tower model is used to obtain the interest preference matrix corresponding to the target user based on the input user data, and the second tower of the two-tower model is used to obtain the interest preference matrix corresponding to the target interest object based on the input recommendation data; The dual-tower model is further provided with an interest matrix layer, in which the interest matrix is ​​pre-configured. The interest matrix layer is connected to the first tower and the second tower respectively, and is used to provide the interest matrix to the first tower and the second tower.

6. The method according to claim 5, wherein the first tower and the second tower further comprise a feature extraction layer and a sub-behavior construction layer, wherein: The feature extraction layer is used to extract features from the attribute information and historical interaction data in the user data or recommendation data of the computing object, respectively, to obtain the object embedding vector of the computing object and the object embedding vectors of each interactive object interacting with the computing object; The sub-behavior construction layer is used to construct a sub-behavior for each interactive object based on the object embedding vector and the object embedding vector of the interactive object.

7. The method according to claim 6, characterized in that The first tower and the second tower further include a multi-interest attention mechanism layer, which is connected to the interest matrix layer and the sub-behavior construction layer respectively; wherein, The multi-interest attention mechanism layer is specifically used to determine the comprehensive preference degree of the computing object for each interest dimension in the interest matrix, and for each sub-behavior, perform feature extraction on the sub-behavior based on the comprehensive preference degree of the computing object for each interest dimension, obtain the feature vector of the computing object in the sub-behavior under each interest dimension, and determine the matrix composed of the feature vectors of each sub-behavior under each interest dimension as the interest preference matrix corresponding to the computing object.

8. A recommendation device, characterized in that: The device includes an acquisition module, a processing module, a calculation module and a judgment module; wherein, The acquisition module is used to acquire user data of a target user and recommendation data of a target object of interest to be recommended to the target user; The acquisition module is further configured to use the target user and the target object of interest as computational objects, respectively, and acquire a sub-behavior set representing all interaction behaviors between the computational object and each interactive object; wherein each interactive object is an interactive object in the interactive object set acquired based on user data of the computational object or historical interaction data in recommendation data; The processing module is used to determine the comprehensive preference degree of the computing object for each interest dimension in the preset interest matrix; The calculation module is used to extract features of each sub-behavior based on the comprehensive preference degree of the calculation object for each interest dimension, obtain the feature vector of the calculation object in each interest dimension in the sub-behavior, and determine the matrix composed of the feature vectors of the calculation object in each interest dimension in each sub-behavior as the interest preference matrix corresponding to the calculation object; The judgment module is configured to determine whether to recommend the target object of interest to the target user based on the interest preference matrix corresponding to the target user and the interest preference matrix corresponding to the target object of interest.

9. A recommendation device, characterized in that The method comprises a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the method according to any one of claims 1 to 7 when executing the program.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images