Media data recall method and device, equipment, storage medium and program product
The user representation model generates behavioral feature unit vectors and media data comparison, and identify high matching data, which solves the problem of low accuracy of traditional recommendation systems and achieves more accurate user interest representation and media data push.
Patent Information
- Application Number
- CN202411099315.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-12
- Publication Date
- 2024-12-06
AI Technical Summary
The traditional recommendation system modeling method is rough and cannot fully reflect the user's interests and needs, resulting in low accuracy of media data push and poor user experience.
The trained user characterization model generates behavioral feature unit vectors of multiple dimensions, compares them with the characteristics of media data, calculates the matching degree between the user and media data, and identifies media data with a matching degree higher than the threshold as recall data.
It improves the accuracy of pushing media data, can more accurately represent user interests, and improves the accuracy and diversity of the recommendation system.
Smart Images

Figure CN119088988A_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the field of computer application technology, and in particular to a method, device, equipment, storage medium and program product for recalling media data. Background Art
[0002] With the popularization of the Internet, media data has experienced explosive growth. Therefore, how to filter out the required media data from the massive media data requires the application of recommendation systems. Recommendation systems are information filtering systems that aim to provide personalized information recommendations based on users' personal preferences and interests. Their core task is to recommend items or services that may be of interest to users by analyzing and predicting their behavior patterns. However, the traditional recommendation system modeling method is rough, and the prediction results are single, which cannot fully reflect the interests and needs of users, resulting in low accuracy in media data push and poor user experience. Based on this, how to improve the accuracy of media data push is a technical problem that needs to be solved urgently. Summary of the invention
[0003] The technical problem to be solved by the embodiments of the present application is to provide a method, apparatus, device, storage medium and program product for recalling media data, which can improve the push accuracy of media data.
[0004] In a first aspect, an embodiment of the present application provides a method for recalling media data, the method comprising:
[0005] Generate a behavior feature unit vector of each dimension based on the behavior feature of the user through an expert network corresponding to each dimension of the multiple dimensions included in the trained user representation model; the behavior feature is obtained based on the interactive behavior data generated during the interaction between the user and the media data;
[0006] For any media data in the media database, the behavior feature unit vectors of each dimension are respectively compared with the media features of the any media data to obtain the matching degree between the media features of the any media data and the behavior feature unit vectors of each dimension;
[0007] Obtaining a matching degree between the user and the any media data according to a matching degree between a media feature of the any media data and a behavior feature unit vector of each dimension;
[0008] The target media data whose matching degree with the user is higher than a preset threshold is identified as the recalled media data of the user.
[0009] In a second aspect, an embodiment of the present application provides a media data recall device, the media data recall device comprising:
[0010] A generating unit, configured to generate a behavior feature unit vector of each dimension based on the behavior feature of the user through an expert network corresponding to each dimension of the multiple dimensions included in the trained user representation model; the behavior feature is obtained based on the interactive behavior data generated during the interaction between the user and the media data;
[0011] A comparison unit, for comparing the behavior feature unit vectors of each dimension with the media features of any media data in the media database, to obtain a matching degree between the media features of any media data and the behavior feature unit vectors of each dimension;
[0012] The generating unit is further configured to obtain a matching degree between the user and the any media data according to a matching degree between a media feature of the any media data and a behavior feature unit vector of each dimension;
[0013] The identification unit is used to identify the target media data whose matching degree with the user is higher than a preset threshold as the recalled media data of the user.
[0014] In one embodiment, the generating unit is further configured to include:
[0015] Generate a shared behavior feature unit vector based on the behavior feature through a shared expert network in the trained user representation model;
[0016] Obtaining a combined behavior feature of any dimension according to the behavior feature unit vector of any dimension and the shared behavior feature unit vector;
[0017] For any media data in the media database, comparing the behavior feature unit vectors of each dimension with the media features of the any media data to obtain the matching degree between the media features of the any media data and the behavior feature unit vectors of each dimension, includes:
[0018] For any media data in the media database, respectively compare the combined behavior features of each dimension with the media features of the any media data to obtain a matching degree between the media features of the any media data and the combined behavior features of each dimension;
[0019] The obtaining the matching degree between the user and the any media data according to the matching degree between the media feature of the any media data and the behavior feature unit vectors of each dimension includes:
[0020] The matching degree between the user and the any media data is obtained according to the matching degree between the media feature of the any media data and the combined behavior features of the various dimensions.
[0021] In one implementation, the generating unit obtains the combined behavior feature of any dimension according to the behavior feature unit vector of any dimension and the shared behavior feature unit vector, including:
[0022] Processing the identity features of the user through the gated network in the trained user representation model to obtain a first weight of the behavior feature unit vector of any dimension and a second weight of the shared behavior feature unit vector, wherein the first weight is used to indicate the contribution degree of the user to any dimension;
[0023] A combined behavior feature of any dimension is obtained according to the behavior feature unit vector of any dimension, the shared behavior feature unit vector, the first weight, and the second weight.
[0024] In one embodiment, the comparison unit is further configured to include:
[0025] Performing dimensionality reduction processing on the behavior feature unit vector of any dimension through the sub-tower corresponding to any dimension in the trained user representation model to obtain the behavior feature unit vector after dimensionality reduction in any dimension;
[0026] For any media data in the media database, comparing the behavior feature unit vectors of each dimension with the media features of the any media data to obtain the matching degree between the media features of the any media data and the behavior feature unit vectors of each dimension, includes:
[0027] For any media data in the media database, the behavior feature unit vectors after dimension reduction of each dimension are respectively compared with the media features of the any media data to obtain the matching degree between the media features of the any media data and the behavior feature unit vectors after dimension reduction of each dimension;
[0028] The obtaining the matching degree between the user and the any media data according to the matching degree between the media feature of the any media data and the behavior feature unit vectors of each dimension includes:
[0029] The matching degree between the user and the any media data is obtained according to the matching degree between the media features of the any media data and the behavior feature unit vectors after dimension reduction in each dimension.
[0030] In one embodiment, the comparison unit is further configured to include:
[0031] splicing the behavior feature unit vector of any dimension and the behavior feature unit vector of any dimension after dimensionality reduction to obtain a spliced behavior feature unit vector of any dimension;
[0032] For any media data in the media database, comparing the reduced-dimensional behavior feature unit vectors of each dimension with the media features of the any media data, respectively, to obtain a matching degree between the media features of the any media data and the reduced-dimensional behavior feature unit vectors of each dimension, includes:
[0033] For any media data in the media database, the spliced behavior feature unit vectors of each dimension are respectively compared with the media features of the any media data to obtain a matching degree between the media features of the any media data and the spliced behavior feature unit vectors of each dimension;
[0034] The obtaining the matching degree between the user and the any media data according to the matching degree between the media feature of the any media data and the behavior feature unit vector after dimension reduction of each dimension includes:
[0035] The matching degree between the user and the any media data is obtained according to the matching degree between the media features of the any media data and the spliced behavior feature unit vectors of the various dimensions.
[0036] In one implementation, the generating unit obtains the matching degree between the user and the any media data according to the matching degree between the media feature of the any media data and the behavior feature unit vectors of each dimension, including:
[0037] According to the weight of each dimension, a weighted sum is performed on the matching degree between the behavior feature unit vector of each dimension and the media feature of any media data to obtain the matching degree between the user and the any media data.
[0038] In one embodiment, the generating unit is further configured to include:
[0039] Generate training behavior feature unit vectors of each dimension based on the behavior features of the training user through an expert network corresponding to each dimension of the multiple dimensions included in the user representation model; the training behavior features are obtained based on the interactive behavior data generated during the interaction between the training user and the training media data;
[0040] For any training media data in the training media database, the training behavior feature unit vectors of the respective dimensions are respectively compared with the media features of the any training media data to obtain a matching degree between the media features of the any training media data and the training behavior feature unit vectors of the respective dimensions;
[0041] Obtaining a matching degree between the training user and the any training media data according to a matching degree between a media feature of the any training media data and a training behavior feature unit vector of each dimension;
[0042] The user representation model is trained according to the matching degree between the training user and any training media data and the preset matching degree to obtain the trained user representation model; the expert network corresponding to each dimension of the multiple dimensions included in the trained user representation model is used to generate the behavior feature unit vector of each dimension based on the user's behavior features.
[0043] In a third aspect, an embodiment of the present application provides a computer device, which includes a memory, a communication interface, and a processor, wherein the memory, the communication interface, and the processor are interconnected; the memory stores a computer program, and the processor calls the computer program stored in the memory to implement the media data recall method of the first aspect above.
[0044] In a fourth aspect, an embodiment of the present application provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, it implements the media data recall method of the first aspect.
[0045] In a fifth aspect, an embodiment of the present application provides a computer program product, which includes a computer program stored in a computer storage medium; a processor of a computer device reads the computer program from the computer storage medium, and the processor executes the computer program, so that the computer device executes the media data recall method of the above-mentioned first aspect.
[0046] In the embodiment of the present application, the expert network corresponding to each dimension in the multiple dimensions included in the trained user representation model generates a behavior feature unit vector of each dimension based on the user's behavior characteristics; the behavior characteristics are obtained based on the interactive behavior data generated during the interaction between the user and the media data; for any media data in the media database, the behavior feature unit vector of each dimension is respectively compared with the media features of any media data to obtain the matching degree between the media features of any media data and the behavior feature unit vector of each dimension; according to the matching degree between the media features of any media data and the behavior feature unit vector of each dimension, the matching degree between the user and any media data is obtained; the target media data whose matching degree with the user is higher than a preset threshold is identified as the user's recalled media data. Therefore, the user's interest points can be accurately represented by generating the behavior feature unit vector of each dimension through the user's behavior characteristics; the matching degree with the media data is calculated according to the behavior feature unit vector of each dimension, and the recalled media data is finally determined, which can improve the push accuracy of the media data. BRIEF DESCRIPTION OF THE DRAWINGS
[0047] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the background technology, the drawings required for use in the embodiments of the present invention or the background technology will be described below.
[0048] Figure 1 is a schematic diagram of the structure of a customized gating network provided in an embodiment of the present application;
[0049] Figure 2 is a schematic diagram of the structure of a gating network provided in an embodiment of the present application;
[0050] Figure 3 is a structural diagram of a user representation model provided in an embodiment of the present application;
[0051] Figure 4 It is a flowchart of a method for training a user representation model provided in an embodiment of the present application;
[0052] Figure 5 It is a flowchart of a method for recalling media data provided by an embodiment of the present application;
[0053] Figure 6 is a structural schematic diagram of a media data recall device provided in an embodiment of the present application;
[0054] Figure 7 It is a structural diagram of a computer device provided in an embodiment of the present application. DETAILED DESCRIPTION
[0055] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.
[0056] It should be noted that, in this article, the terms "include", "comprises" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also includes other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, an element defined by the sentence "includes a ..." does not exclude the existence of other identical elements in the process, method, article or device including the element. In addition, components, features, and elements with the same name in different embodiments of the present application may have the same meaning or different meanings, and their specific meanings need to be determined by their explanation in the specific embodiment or further combined with the context of the specific embodiment.
[0057] At the same time, in the description of the embodiments of the present application, the terms "first", "second", etc. are only used to distinguish the description and cannot be understood as indicating or implying relative importance. Therefore, the features defined as "first" and "second" may explicitly or implicitly include one or more features. In the description of the embodiments of the present application, the meaning of "multiple" is two or more, unless otherwise clearly and specifically defined.
[0058] In the specific implementation of this application, user-related data, such as user behavior data, is involved. When the embodiments of this application are applied to specific products or technologies, user permission or consent is required, and the collection, use and processing of relevant data must comply with local laws, regulations and standards.
[0059] The media data recall method provided in the embodiment of the present application is applied in the content recommendation system, and can recall the content (media data) of interest to the user based on the user behavior data, where the user can also be referred to as an object. Recommendation systems usually focus on multiple types of goals. For example, improving clicks and interaction rates will increase the degree of match between content and user interests, which helps to form conversions; improving viewing time is conducive to increasing user stickiness and helping user retention. The embodiment of the present application improves the structure of the dual-tower model and optimizes multiple goals such as clicks and duration at the same time, thereby improving the accuracy and diversity of the recommendation system.
[0060] Among them, the dual-tower model is used to solve the personalized recommendation problem in the recommendation system. The model structure is based on two parallel neural networks (towers). Each tower processes the information of users and items (or content) respectively, and then combines them to make the final recommendation decision. User Tower: The input of the user tower is information about the user, such as the user's ID, historical behavior, personal data, etc. The goal of the user tower is to learn the user's feature representation and map the user's features to a low-dimensional continuous vector space; Item Tower: The input of the item tower is information about the item (or content), such as the item's ID, attributes, text description, etc. The goal of the item tower is to learn the feature representation of the item and map the item's features to the same low-dimensional continuous vector space. The embodiment of the present application still adopts a dual-tower structure, keeping the model structure of the item (note) side unchanged, and the user-side model adopts a customized gate control network (Customized Gate Control, CGC) structure. Each target has its own corresponding expert network, as well as a shared expert network. After CGC, multiple vectors of the user are obtained, corresponding to different optimization targets.
[0061] CGC refers to a model architecture in deep learning that allows users to customize the gating mechanism of neural networks according to specific tasks and needs. CGC draws on the idea of Multi-gate Mixture-of-Experts. Its main structure includes input layer, hidden layer, gating network layer and output layer. Figure 1 , Figure 1 It is a structural diagram of a customized gated network provided in an embodiment of the present application. At the top are some multi-layer tower networks related to the task, such as Tower A, i.e. sub-tower A, and output A of sub-tower A, Tower B, i.e. sub-tower B, and output B of sub-tower B. At the bottom are some expert modules, each of which is composed of multiple expert networks. Expert modules are divided into two categories, one is the task-shared expert module, which is responsible for learning the shared mode of the task, for example, Experts Shared, i.e. shared expert network; the other is the task-exclusive expert module, which is responsible for learning the exclusive mode of the task, for example, Experts A, i.e. expert network A, and Experts B, i.e. expert network B. Each tower network learns knowledge from all expert modules (shared expert modules and exclusive expert modules). For example, Experts A (expert network A) includes expert network E. A,1 , E A,2 …E A,mA , Experts B (expert network B) includes expert network E B,1 , E B,2 …E B,mB , Experts Shared (shared expert network) includes expert network ES,1 , E S,2 …E S,mS The input of each sub-tower is the feature representation of each expert network and the shared expert network fused by CGC, and then converted into a dense embedding representation of the same dimension as the output of the note side through a multi-layer neural network. For example, T in Tower A A,1 …T A,nA and the T in Tower B B,1 …T B,nB .
[0062] In CGC, the outputs of all expert modules are fused through a gating network. Figure 2 , Figure 2 It is a structural diagram of a gating network provided in an embodiment of the present application. The gating network is a single-layer feedforward network that uses softmax as an activation function. The input is used as a selector to calculate the weighted sum of all selected vectors. That is, the selection probability is generated based on the input selector, and then the outputs Vector 1, ..., Vector m of all expert modules are fused based on the selection probability, so the input is called a selector; the selected vectors are the outputs of the expert modules, which are selected based on the selection probability. Among them, FC refers to the fully connected layer (Fully Connected, FC). In the neural network, each neuron in the FC layer is connected to each neuron in the previous layer, so that each neuron can receive the outputs of all neurons in the previous layer, and can be weighted by weights.
[0063] In the embodiment of the present application, the user side model, i.e., the user representation model, adopts a customized gated network structure. The user representation model is also called an object representation model. Figure 3 , Figure 3 It is a structural diagram of a user characterization model provided in an embodiment of the present application. The main structure of the user characterization model includes an input layer, an embedding layer, a hidden layer, a gated network layer and an output layer.
[0064] The input layer inputs the user's identity features and user behavior sequence. The user behavior sequence is obtained based on the interactive behavior data generated during the user's interaction with media data, and the user's identity features are obtained based on the user's basic information. The interactive behavior data indicates the user's interest points. The interactive behavior data may include data on click operations performed by the user's corresponding account, data on the browsing time of the user's corresponding account exceeding a preset threshold, and data on interactive operations performed by the user's corresponding account such as likes, favorites, forwarding or comments. The user's basic information indicates the user's type, such as male or female, white-collar or student, consumption level, etc. Users of the same type may have the same interests or similar behaviors. The user's basic information data may include the user's age, occupation, gender, social account, address, etc. uploaded when registering an account, the login time, online time, login frequency, consumption amount, etc. of the user's corresponding account.
[0065] Among them, the embedding layer is used to map discrete, high-dimensional input data into a continuous, low-dimensional vector space, that is, to convert the input user's identity features and user behavior sequences into dense representation results that can be used by the upper-level model.
[0066] The hidden layer uses multiple expert networks, including a shared expert network and three unique expert networks, namely the click expert network, the interaction expert network and the duration expert network. Each expert network is a fully connected layer, which is used to learn the different representations and shared representations of the input user behavior characteristics corresponding to each dimension, so as to learn and transfer the commonalities between targets. The click expert network corresponds to the click dimension, and the click expert network is used to generate the behavior feature unit vector of the click dimension based on the user's behavior characteristics. The click expert network includes the expert network E c,1 , expert network E c,2 , expert network E c,3 The interactive expert network corresponds to the interactive dimension. The interactive expert network is used to generate the behavioral feature unit vector of the interactive dimension based on the user's behavioral features. The interactive expert network includes the expert network E e,1 , expert network E e,2 , expert network E e,3 The time-length expert network corresponds to the time-length dimension. The time-length expert network is used to generate a behavior feature unit vector of the time-length dimension based on the user's behavior characteristics. The time-length expert network includes the expert network E t,1 , expert network E t,2 , expert network E t,3 The shared expert network is used to generate a shared behavior feature unit vector based on the user's behavior features. The shared expert network includes an expert network E s,1 , expert network E s,2 , expert network Es,3 .
[0067] In the gated network layer, each dimension has its own independent gated network. Click the Gate corresponding to the expert network. c Gated network c, Gate corresponding to the interactive expert network e Gated network e, Gate corresponding to the duration expert network t Gated network t. Each gating network is used to calculate the importance weight of each expert network for the current dimension. These weights will be used to perform weighted summation on the outputs of each expert network to obtain the feature representation output of the dimension. The gating network uses identity features as feature input to learn the common interests of users of the same type. For example, since the item side (note side) can include text or video notes, different types of users have different preferences for text or video notes. Young people aged 20-40 may prefer text notes, and older or younger people may prefer video notes. The gating network can control the vector output by the user representation model according to different user types and is related to the user type preference. The output formula of the user representation model is as follows (1):
[0068] g k (x) = w k (x)S k (x) (1)
[0069] Among them, x represents user features, which include user identity features and user behavior sequences, k represents the kth dimension, and w k (x) is the weight vector of the expert network used in dimension k, S k (x) is the matrix of the expert network used by dimension k. k (x) is the sum vector of the products of each element of the weight vector and the rows of the matrix. k The formula of (x) is as follows:
[0070]
[0071] in, Represents the fully connected parameter matrix of the softmax of dimension k.
[0072] S k The formula of (x) is as follows:
[0073]
[0074] Among them, E represents the expert network, represents the transpose of the output features of the first expert of dimension k, m k is the number of expert networks for task k, m sis the number of expert networks shared by all tasks.
[0075] Among them, the output layer corresponds to the click sub-tower, the interaction sub-tower and the duration sub-tower respectively. The output of the click sub-tower is the user click output, that is, the click feature unit vector; the output of the interaction sub-tower is the user interaction output, that is, the interaction feature unit vector; the output of the duration sub-tower is the user duration output, that is, the duration feature unit vector; the input of each sub-tower is the feature representation g of the expert network and the shared expert network corresponding to each dimension after CGC fusion. k (x), after passing through a multi-layer neural network, is converted into a dense embedding representation of the same dimension as the output on the note side, for example, click on the T in the sub-tower c,1 , T c,... , T in the interactive subtower e,1 , T e,... And the T in the long sub-tower t,1 , T t,... Among them, the output of CGC will be additionally spliced as a residual to the top layer of each dimensional sub-tower, which can reduce the gradient loss of the user representation model, speed up the update speed of the user representation model, and reduce information loss.
[0076] Based on the above description, see Figure 4 , Figure 4 is a flow chart of a method for training a user representation model provided in an embodiment of the present application, such as Figure 4 The training method of the user representation model shown includes but is not limited to steps S401 to S404, wherein:
[0077] S401, generating a training behavior feature unit vector of each dimension based on the behavior features of the training user through an expert network corresponding to each dimension in the multiple dimensions included in the user representation model.
[0078] Among them, users can also be called objects, user representation models can also be called object representation models, and training users can also be called training objects. Multiple dimensions can include clicks, interactions, and duration, which correspond to click expert networks, interactive expert networks, and duration expert networks, respectively. Interaction dimensions can include likes, favorites, comments, and reposts. The training behavior feature unit vector of the training user is obtained based on the interactive behavior data generated during the interaction between the training user and the media data. Each expert network is a fully connected layer, which is used to learn to extract different representations of each dimension corresponding to the input feature. The expert network can be a deep neural network.
[0079] Among them, the interactive behavior data of training users may include data generated by training users clicking on media data, data generated by training users browsing media data for a preset time threshold, and data on the interaction between training users and media data, such as data generated by training users liking media data, data generated by training users collecting media data, data generated by training users commenting on media data, and data generated by training users forwarding media data. Specifically, the data generated by training users clicking on media data may include the training user's click information, the media data's identification, type, account information for uploading media data, content information of media data, etc.; the data generated by training users browsing media data for a preset time threshold may include the training user's browsing time, the media data's identification, type, account information for uploading media data, content information of media data, etc.; the data generated by training users liking media data may include the training user's like information, the media data's identification, type, account information for uploading media data, content information of media data, etc.; the data generated by training users collecting media data may include the training user's collection information, the media data's identification, type, account information for uploading media data, content information of media data, etc.; the data generated by training users commenting on media data may include the training user's comment information, the media data's identification, type, account information for uploading media data, content information of media data, etc.; the data generated by training users forwarding media data may include the training user's forwarding information, the training user's forwarding object information, the media data's identification, type, account information for uploading media data, content information of media data, etc.
[0080] Optionally, the training of expert networks (deep neural networks) of various dimensions usually uses a back-propagation algorithm. The algorithm calculates the error between the predicted output and the actual output, and back-propagates the error to update the weights and bias parameters in the network. Through repeated iterative optimization, the network can adjust parameters to minimize the prediction error and gradually improve the performance of the model. Therefore, after the training is completed, the expert network of any dimension can generate a training behavior feature unit vector of that dimension.
[0081] In an optional embodiment, the following methods can also be used: generating a shared training behavior feature unit vector based on the training behavior feature through a shared expert network in the user characterization model; obtaining a combined training behavior feature of any dimension according to the training behavior feature unit vector and the shared training behavior feature unit vector of any dimension. Among them, the generation of the combined training behavior feature can be done by any of the following methods: feature concatenation: concatenating the training behavior feature unit vector and the shared training behavior feature unit vector of any dimension to form a longer vector; weighted operation: weights can be used to perform weighted operations on the training behavior feature unit vector and the shared training behavior feature unit vector of any dimension to adjust the contribution of the training behavior feature unit vector and the shared training behavior feature unit vector of any dimension to the combined behavior feature; feature interaction: if there is a clear interactive relationship between the training behavior feature unit vector and the shared training behavior feature unit vector of any dimension, the combined training behavior feature can be generated by element product or other nonlinear operations.
[0082] Among them, the shared expert network learns the commonality of the training behavior features according to the input training behavior features to output the shared training behavior feature unit vector of the shared representation. Optionally, the network can be trained using training data to minimize the loss of shared features, the gradient is calculated using the back propagation algorithm, and the network parameters are updated by the optimization algorithm (such as Adam, SGD, etc.), and the shared expert network is obtained after the training is completed. Specifically, the training data includes data of multiple dimensions (clicks, interactions, duration), and the training data is input into the network so that the hidden layer in the network extracts the shared features of the training data, wherein the shared features of the training data include the shared representation of multiple dimensional features, and the difference between the extracted shared features and the preset shared features, i.e., the loss value, is calculated by the preset loss function, wherein the preset function may include the mean square error function and the contrast loss function, etc., and the network is trained in the direction of reducing the loss value, the gradient is calculated using the back propagation algorithm, and the network parameters are updated by the optimization algorithm (such as Adam, SGD, etc.) to minimize the loss value, and the shared expert network is obtained after the training is completed. The shared expert network obtained by such training can accurately extract the shared features of multiple dimensions to output the shared training behavior feature unit vector of the shared representation.
[0083] In an optional embodiment, based on the training behavior feature unit vector and the shared training behavior feature unit vector of any dimension, a combined training behavior feature of any dimension is obtained. The method can be as follows: the identity characteristics of the training user are processed through a gating network in a user representation model to obtain a first weight of the training behavior feature unit vector of the training user in any dimension and a second weight of the shared training behavior feature unit vector, wherein the first weight is used to indicate the degree of contribution of the training user to any dimension; based on the training behavior feature unit vector, the shared training behavior feature unit vector, the first weight and the second weight of any dimension, a combined training behavior feature of any dimension is obtained.
[0084] Among them, the identity characteristics of the training user are obtained based on the basic information of the training user. The basic information of the training user indicates the type of the training user, such as male or female, white-collar worker or student, consumption level, etc. The same type of training users may have the same interests or similar behaviors. The basic information data of the training user may include the age, occupation, gender, social account, address and other information uploaded by the training user when registering an account, the login time, online time, login frequency, consumption amount and other information of the account corresponding to the training user. The gated network can be a fully connected network, and its purpose is to extract weights related to any dimension from the user identity characteristics. The output of the gated network is the first weight of the training behavior feature unit vector of any dimension and the second weight of the shared training behavior feature unit vector. The first weight is used to measure the importance or contribution of the user in this dimension, and the first weight and the second weight add up to 1. According to the training behavior feature unit vector of any dimension, the shared training behavior feature unit vector, the first weight and the second weight, a weighted operation is performed to obtain the combined training behavior feature of any dimension.
[0085] In an optional embodiment, the following method can be used: the training behavior feature unit vector of any dimension is subjected to dimensionality reduction processing through the sub-tower corresponding to any dimension in the user characterization model to obtain the reduced-dimensional training behavior feature unit vector of any dimension. The purpose of dimensionality reduction is to reduce the dimension of the data through mathematical transformation while retaining as much information as possible, such as principal component analysis, linear discriminant analysis, etc. After the output of the sub-tower is subjected to dimensionality reduction processing, the result is a more compact vector with fewer dimensions, that is, the reduced-dimensional training behavior feature unit vector. By using a sub-tower to perform dimensionality reduction processing on the training behavior feature unit vector of a specific dimension, the complexity of the user characterization model can be effectively simplified and its efficiency and interpretability can be improved. The purpose of dimensionality reduction also includes making the reduced-dimensional training behavior feature unit vector have the same dimension as the note feature output on the note side.
[0086] In an optional embodiment, the following method can also be used: splicing the training behavior feature unit vector of any dimension and the training behavior feature unit vector after dimensionality reduction of any dimension to obtain the spliced training behavior feature unit vector of any dimension. Among them, the splicing can be done in the following ways: simple splicing: connecting the training behavior feature unit vector of any dimension and the training behavior feature unit vector after dimensionality reduction of any dimension in sequence to obtain the spliced training behavior feature unit vector of any dimension. Weighted splicing: weighted combination of the training behavior feature unit vector of any dimension and the training behavior feature unit vector after dimensionality reduction of any dimension to balance the contribution of the training behavior feature unit vector of any dimension and the training behavior feature unit vector after dimensionality reduction of any dimension in the spliced training behavior feature unit vector of any dimension. By additionally splicing the output of CGC as a residual to the top layer of each subtower, information loss can be reduced.
[0087] S402, for any training media data in the training media database, respectively compare the training behavior feature unit vectors of each dimension with the media features of any training media data to obtain the matching degree between the media features of any training media data and the training behavior feature unit vectors of each dimension.
[0088] Among them, each training media data has its corresponding media features, and the media features include data about the media content, such as the length, label, and number of views of the video. Different similarity metrics can be used to evaluate the match between the training behavior feature unit vector and the media features. For example, the following methods can be used: cosine similarity: measures the angle between vectors, and the closer the value is to 1, the more similar it is. Euclidean distance: measures the spatial distance between vectors, and the smaller the distance, the more similar it is. Pearson correlation coefficient: measures the strength and direction of linear correlation, ranging from -1 to 1.
[0089] In an optional implementation, for any training media data in the training media database, the training behavior feature unit vectors of each dimension are respectively compared with the media features of any training media data to obtain the matching degree between the media features of any training media data and the training behavior feature unit vectors of each dimension. This can be done in the following way: for any training media data in the training media database, the combined training behavior features of each dimension are respectively compared with the media features of any training media data to obtain the matching degree between the media features of any training media data and the combined training behavior features of each dimension.
[0090] In an optional implementation, for any training media data in the training media database, the training behavior feature unit vectors of each dimension are respectively compared with the media features of any training media data to obtain the matching degree between the media features of any training media data and the training behavior feature unit vectors of each dimension. This can be done in the following way: for any training media data in the training media database, the training behavior feature unit vectors of each dimension after dimensionality reduction are respectively compared with the media features of any training media data to obtain the matching degree between the media features of any training media data and the training behavior feature unit vectors of each dimension after dimensionality reduction.
[0091] In an optional implementation, for any training media data in the training media database, the reduced-dimensional training behavior feature unit vectors of each dimension are respectively compared with the media features of any training media data to obtain the matching degree between the media features of any training media data and the reduced-dimensional training behavior feature unit vectors of each dimension. This can be done in the following way: for any training media data in the training media database, the spliced training behavior feature unit vectors of each dimension are respectively compared with the media features of any training media data to obtain the matching degree between the media features of any training media data and the spliced training behavior feature unit vectors of each dimension.
[0092] S403, obtaining a matching degree between a training user and any training media data according to a matching degree between a media feature of any training media data and a training behavior feature unit vector of each dimension.
[0093] After obtaining multiple matching scores for each dimension, these scores can be combined to evaluate the overall matching degree.
[0094] In an optional implementation, the matching degree between the training user and any training media data is obtained based on the matching degree between the media features of any training media data and the training behavior feature unit vectors of each dimension. The matching degree between the training user and any training media data can be obtained in the following manner: according to the weights of each dimension, the matching degrees between the training behavior feature unit vectors of each dimension and the media features of any training media data are weighted and summed to obtain the matching degree between the training user and any training media data. The weights of the training behavior feature unit vectors of each dimension are obtained based on data analysis or model training, and the weighted summation of each training behavior feature unit vector is performed to obtain the matching degree between the training user and any training media data.
[0095] In an optional embodiment, the matching degree between the training user and any training media data is obtained based on the matching degree between the media features of any training media data and the training behavior feature unit vectors of each dimension. This can be done in the following way: the matching degree between the training user and any training media data is obtained based on the matching degree between the media features of any training media data and the combined training behavior features of each dimension.
[0096] In an optional embodiment, the matching degree between the training user and any training media data is obtained based on the matching degree between the media features of any training media data and the training behavior feature unit vectors of each dimension. This can be done in the following way: the matching degree between the training user and any training media data is obtained based on the matching degree between the media features of any training media data and the reduced-dimensional training behavior feature unit vectors of each dimension.
[0097] In an optional embodiment, the matching degree between the training user and any training media data is obtained based on the matching degree between the media features of any training media data and the training behavior feature unit vectors after dimensionality reduction in each dimension. This can be done in the following way: the matching degree between the training user and any training media data is obtained based on the matching degree between the media features of any training media data and the spliced training behavior feature unit vectors in each dimension.
[0098] S404: Train the user representation model according to the matching degree between the training user and any training media data and the preset matching degree to obtain a trained user representation model.
[0099] Among them, the loss value is calculated by the matching degree between the training user and any training media data and the preset matching degree, and the user representation model is trained using the back propagation algorithm and the optimizer. During the training process, the user representation model will continuously adjust the parameters to minimize the loss value, and finally obtain the trained user representation model. The expert network corresponding to each of the multiple dimensions contained in the trained user representation model is used to generate the behavior feature unit vector of each dimension based on the user's behavior characteristics.
[0100] In an optional implementation, the user representation model is trained based on the matching degree between the media features of any training media data and the training behavior feature unit vectors of each dimension, and the preset matching degree corresponding to each dimension, to obtain a trained user representation model. The loss value is calculated by respectively comparing the training behavior feature unit vectors obtained through the sub-towers of each dimension with the preset matching degree corresponding to each dimension, and the loss weight is set according to the importance between the dimensions, and the user representation model is trained to obtain a trained user representation model.
[0101] In the embodiment of the present application, the training behavior feature unit vector of each dimension is generated based on the behavior features of the training user through the expert network corresponding to each dimension in the multiple dimensions included in the user characterization model; for any training media data in the training media database, the training behavior feature unit vector of each dimension is respectively compared with the media features of any training media data to obtain the matching degree between the media features of any training media data and the training behavior feature unit vector of each dimension; according to the matching degree between the media features of any training media data and the training behavior feature unit vector of each dimension, the matching degree between the training user and any training media data is obtained; according to the matching degree between the training user and any training media data and the preset matching degree, the user characterization model is trained to obtain the trained user characterization model. Therefore, the training behavior feature unit vector of each dimension is generated through the training behavior features of the training user; the matching degree with the training media data is calculated according to the training behavior feature unit vector of each dimension, and the user characterization model is trained according to the matching degree and the preset matching degree to obtain the trained user characterization model, and the trained user characterization model can accurately characterize the interest points of the training user.
[0102] Based on the above description, see Figure 5 , Figure 5 is a flow chart of a method for recalling media data provided by an embodiment of the present application, such as Figure 5 The media data recall method shown includes but is not limited to steps S501 to S504, wherein:
[0103] S501, generating a behavior feature unit vector of each dimension based on the behavior features of the user through an expert network corresponding to each dimension in the multiple dimensions included in the trained user representation model.
[0104] Among them, users can also be called objects, and user representation models can also be called object representation models. Multiple dimensions can include clicks, interactions, and duration, which correspond to click expert networks, interactive expert networks, and duration expert networks, respectively. Interaction dimensions can include likes, favorites, comments, and reposts. The user's behavioral feature unit vector is obtained based on the interactive behavior data generated during the user's interaction with media data. Each expert network is a fully connected layer used to learn and extract different representations of input features corresponding to each dimension. The expert network can be a deep neural network.
[0105] Among them, the user's interactive behavior data may include data generated by the user clicking on media data, data generated by the user browsing media data for a preset time threshold, and data on the user's interaction with media data, such as data generated by the user liking media data, data generated by the user collecting media data, data generated by the user commenting on media data, and data generated by the user forwarding media data. Specifically, the data generated by a user clicking on media data may include the user's click information, the media data's identification, type, account information for uploading the media data, content information of the media data, etc.; the data generated by a user browsing media data for a preset time threshold may include the user's browsing time, the media data's identification, type, account information for uploading the media data, content information of the media data, etc.; the data generated by a user liking media data may include the user's like information, the media data's identification, type, account information for uploading the media data, content information of the media data, etc.; the data generated by a user collecting media data may include the user's collection information, the media data's identification, type, account information for uploading the media data, content information of the media data, etc.; the data generated by a user commenting on media data may include the user's comment information, the media data's identification, type, account information for uploading the media data, content information of the media data, etc.; the data generated by a user forwarding media data may include the user's forwarding information, the user's forwarding object information, the media data's identification, type, account information for uploading the media data, content information of the media data, etc.
[0106] In an optional embodiment, the following method can also be used: a shared behavior feature unit vector is generated based on the behavior feature through a shared expert network in a trained user representation model; a combined behavior feature of any dimension is obtained according to the behavior feature unit vector of any dimension and the shared behavior feature unit vector. The combined behavior feature can be generated by any of the following methods: feature concatenation: the behavior feature unit vector of any dimension and the shared behavior feature unit vector are concatenated together to form a longer vector; weighted operation: weights can be used to perform weighted operations on the behavior feature unit vector and the shared behavior feature unit vector of any dimension to adjust the contribution of the behavior feature unit vector and the shared behavior feature unit vector of any dimension to the combined behavior feature; feature interaction: if there is a clear interactive relationship between the behavior feature unit vector and the shared behavior feature unit vector of any dimension, the combined behavior feature can be generated by element multiplication or other nonlinear operations. The shared expert network learns the commonality of the behavior features based on the input behavior features to output the shared behavior feature unit vector of the shared representation.
[0107] In an optional embodiment, based on the behavior feature unit vector and the shared behavior feature unit vector of any dimension, a combined behavior feature of any dimension is obtained. This can be done in the following way: the identity features of the user are processed by a gating network in a trained user representation model to obtain a first weight of the behavior feature unit vector of the user in any dimension and a second weight of the shared behavior feature unit vector, wherein the first weight is used to indicate the degree of contribution of the user to any dimension; based on the behavior feature unit vector, the shared behavior feature unit vector, the first weight and the second weight of any dimension, a combined behavior feature of any dimension is obtained.
[0108] Among them, the identity characteristics of the user are obtained based on the basic information of the user. The basic information of the user indicates the type of user, such as male or female, white-collar worker or student, consumption level, etc. Users of the same type may have the same interests or similar behaviors. The basic information data of the user may include the age, occupation, gender, social account, address and other information uploaded by the user when registering the account, the login time, online time, login frequency, consumption amount and other information of the account corresponding to the user. The gated network can be a fully connected network, and its purpose is to extract weights related to any dimension from the user's identity characteristics. The output of the gated network is the first weight of the behavior feature unit vector of any dimension and the second weight of the shared behavior feature unit vector. The first weight is used to measure the importance or contribution of the user in this dimension, and the first weight and the second weight add up to 1. According to the behavior feature unit vector of any dimension, the shared behavior feature unit vector, the first weight and the second weight, a weighted operation is performed to obtain the combined behavior feature of any dimension.
[0109] In an optional embodiment, the following method can be used: the behavior feature unit vector of any dimension is subjected to dimensionality reduction processing through the sub-tower corresponding to any dimension in the trained user representation model to obtain the reduced-dimensional behavior feature unit vector of any dimension. The purpose of dimensionality reduction is to reduce the dimension of the data through mathematical transformation while retaining as much information as possible, such as principal component analysis, linear discriminant analysis, etc. After the output of the sub-tower is subjected to dimensionality reduction processing, the result is a more compact vector with fewer dimensions, that is, the reduced-dimensional behavior feature unit vector. By using a sub-tower to perform dimensionality reduction processing on the behavior feature unit vector of a specific dimension, the complexity of the trained user representation model can be effectively simplified and its efficiency and interpretability can be improved. The purpose of dimensionality reduction also includes making the reduced-dimensional behavior feature unit vector the same dimension as the note feature output by the note side.
[0110] In an optional embodiment, the following method can also be used: splicing the behavior feature unit vector of any dimension and the behavior feature unit vector after dimensionality reduction of any dimension to obtain the spliced behavior feature unit vector of any dimension. Among them, the splicing can be done in the following ways: simple splicing: connecting the behavior feature unit vector of any dimension and the behavior feature unit vector after dimensionality reduction of any dimension in sequence to obtain the spliced behavior feature unit vector of any dimension. Weighted splicing: weighted combination of the behavior feature unit vector of any dimension and the behavior feature unit vector after dimensionality reduction of any dimension to balance the contribution of the behavior feature unit vector of any dimension and the behavior feature unit vector after dimensionality reduction of any dimension in the spliced behavior feature unit vector of any dimension. By additionally splicing the output of CGC as a residual to the top layer of each subtower, information loss can be reduced.
[0111] S502 , for any media data in the media database, respectively compare the behavior feature unit vectors of each dimension with the media features of any media data to obtain the matching degree between the media features of any media data and the behavior feature unit vectors of each dimension.
[0112] Among them, each media data has its corresponding media feature, which includes data about the media content, such as the length, label, and number of views of the video. Different similarity metrics can be used to evaluate the match between the training behavior feature unit vector and the media features. For example, the following methods can be used: cosine similarity: measures the angle between vectors, the closer the value is to 1, the more similar it is. Euclidean distance: measures the spatial distance between vectors, the smaller the distance, the more similar it is. Pearson correlation coefficient: measures the strength and direction of linear correlation, ranging from -1 to 1.
[0113] In an optional implementation, the media features of each media data in the media database are extracted through the content representation model, and the media features are stored in the media feature library. The media data can indicate notes, and the note-side vector is calculated and stored in advance. When the behavioral feature unit vectors of each dimension are compared with the media features of any media data, the media features are directly obtained from the media feature library, which can improve efficiency and save resources.
[0114] In an optional implementation, for any media data in a media database, the behavioral feature unit vectors of each dimension are respectively compared with the media features of any media data to obtain the matching degree between the media features of any media data and the behavioral feature unit vectors of each dimension. This can be done in the following way: for any media data in a media database, the combined behavioral features of each dimension are respectively compared with the media features of any media data to obtain the matching degree between the media features of any media data and the combined behavioral features of each dimension.
[0115] In an optional implementation, for any media data in a media database, the behavior feature unit vectors of each dimension are respectively compared with the media features of any media data to obtain the matching degree between the media features of any media data and the behavior feature unit vectors of each dimension. This can be done in the following way: for any media data in a media database, the behavior feature unit vectors of each dimension after dimensionality reduction are respectively compared with the media features of any media data to obtain the matching degree between the media features of any media data and the behavior feature unit vectors of each dimension after dimensionality reduction.
[0116] In an optional implementation, for any media data in a media database, the reduced-dimensional behavior feature unit vectors of each dimension are respectively compared with the media features of any media data to obtain the matching degree between the media features of any media data and the reduced-dimensional behavior feature unit vectors of each dimension. This can be done in the following way: for any media data in a media database, the spliced behavior feature unit vectors of each dimension are respectively compared with the media features of any media data to obtain the matching degree between the media features of any media data and the spliced behavior feature unit vectors of each dimension.
[0117] S503: Obtain the matching degree between the user and any media data according to the matching degree between the media features of any media data and the behavior feature unit vectors of each dimension.
[0118] After obtaining multiple matching scores for each dimension, these scores can be combined to evaluate the overall matching degree.
[0119] In an optional implementation, the degree of matching between a user and any media data is obtained based on the degree of matching between the media features of any media data and the behavioral feature unit vectors of each dimension, which can be done in the following way: according to the weights of each dimension, the degree of matching between the behavioral feature unit vectors of each dimension and the media features of any media data is weighted and summed to obtain the degree of matching between the user and any media data. The weights of the behavioral feature unit vectors of each dimension can be obtained based on data analysis or model training, and the weighted summation of each behavioral feature unit vector is performed to obtain the degree of matching between the user and any media data.
[0120] In an optional implementation, the matching degree between the user and any media data is obtained based on the matching degree between the media features of any media data and the behavioral feature unit vectors of each dimension. This can be done in the following way: the matching degree between the user and any media data is obtained based on the matching degree between the media features of any media data and the combined behavioral features of each dimension.
[0121] In an optional implementation, the degree of matching between a user and any media data is obtained based on the degree of matching between the media features of any media data and the behavioral feature unit vectors of each dimension. This can be done in the following way: the degree of matching between a user and any media data is obtained based on the degree of matching between the media features of any media data and the reduced-dimensional behavioral feature unit vectors of each dimension.
[0122] In an optional implementation, the degree of matching between a user and any media data is obtained based on the degree of matching between the media features of any media data and the behavioral feature unit vectors after dimensionality reduction of each dimension. This can be done in the following way: the degree of matching between a user and any media data is obtained based on the degree of matching between the media features of any media data and the spliced behavioral feature unit vectors of each dimension.
[0123] S504: Identify the target media data whose matching degree with the user is higher than a preset threshold as the user's recalled media data.
[0124] If the matching degree is higher than a preset threshold, the media data is the user's recalled media data; if the matching degree is lower than or equal to the preset threshold, the media data is discarded.
[0125] Optionally, when a user requests a recommendation online, multiple features of the user are obtained in real time. These features correspond to the user's different interests, behavior patterns, or other personalized information, such as viewing history, preference tags, etc. Each user feature is inferred in real time through an online model to capture the user's current interests and features. For each user feature, the ANN (Approximate Nearest Neighbor) algorithm is used to find a set of notes that are most similar to it in the pre-calculated note-side features. The ANN algorithm can efficiently find approximate nearest neighbors in large data sets, thereby quickly filtering out notes related to the user's current interests. For each set of notes retrieved by each user feature, subsequent sorting processing is performed. After the sorting stage, a portion of the top-ranked notes are displayed to the user as the result of personalized recommendation.
[0126] Optionally, the matching degree can be combined with other features of the media data to obtain a sorting result. For example, the popularity, importance, timeliness, etc. of the notes can be combined to weight these features and obtain the contribution of these features in the recommendation sorting. Use a sorting algorithm to sort the weighted scores. The sorting algorithm can be used in the following ways: Rule-based sorting: Sort notes according to preset rules (such as priority, timeliness, etc.). Machine learning-based sorting: Use machine learning models (such as ranking neural networks, sorting models) to predict user clicks or interaction behaviors and sort them accordingly. Traditional sorting algorithms: For example, score-based sorting (sorting from high to low by score) or text similarity-based sorting (sorting from high to low by similarity).
[0127] In the embodiment of the present application, the expert network corresponding to each dimension in the multiple dimensions included in the trained user representation model generates a behavior feature unit vector of each dimension based on the user's behavior characteristics; for any media data in the media database, the behavior feature unit vector of each dimension is respectively compared with the media features of any media data to obtain the matching degree between the media features of any media data and the behavior feature unit vector of each dimension; according to the matching degree between the media features of any media data and the behavior feature unit vector of each dimension, the matching degree between the user and any media data is obtained; the target media data whose matching degree with the user is higher than a preset threshold is identified as the user's recalled media data. Therefore, generating the behavior feature unit vector of each dimension based on the user's behavior characteristics can accurately represent the user's interest points; calculating the matching degree with the media data according to the behavior feature unit vector of each dimension, and finally determining the recalled media data, can improve the push accuracy of the media data.
[0128] An embodiment of the present application further provides a computer storage medium, in which program instructions are stored. When the program instructions are executed, they are used to implement the corresponding methods described in the above embodiments.
[0129] See also Figure 6 , Figure 6 It is a structural schematic diagram of a media data recall device provided in an embodiment of the present application.
[0130] In one implementation of the media data recall device of the embodiment of the present application, the media data recall device includes the following structure:
[0131] A generating unit 601 is used to generate a behavior feature unit vector of each dimension based on the behavior feature of the user through an expert network corresponding to each dimension of the multiple dimensions included in the trained user representation model; the behavior feature is obtained based on the interactive behavior data generated during the interaction between the user and the media data;
[0132] The comparison unit 602 is used to compare the behavior feature unit vectors of each dimension with the media features of any media data in the media database, and obtain the matching degree between the media features of any media data and the behavior feature unit vectors of each dimension;
[0133] The generating unit 601 is further configured to obtain a matching degree between the user and the any media data according to a matching degree between a media feature of the any media data and a behavior feature unit vector of each dimension;
[0134] The identification unit 603 is configured to identify the target media data whose matching degree with the user is higher than a preset threshold as the recalled media data of the user.
[0135] In one implementation, the generating unit 601 is further configured to include:
[0136] Generate a shared behavior feature unit vector based on the behavior feature through a shared expert network in the trained user representation model;
[0137] Obtaining a combined behavior feature of any dimension according to the behavior feature unit vector of any dimension and the shared behavior feature unit vector;
[0138] For any media data in the media database, comparing the behavior feature unit vectors of each dimension with the media features of the any media data to obtain the matching degree between the media features of the any media data and the behavior feature unit vectors of each dimension, includes:
[0139] For any media data in the media database, respectively compare the combined behavior features of each dimension with the media features of the any media data to obtain a matching degree between the media features of the any media data and the combined behavior features of each dimension;
[0140] The obtaining the matching degree between the user and the any media data according to the matching degree between the media feature of the any media data and the behavior feature unit vectors of each dimension includes:
[0141] The matching degree between the user and the any media data is obtained according to the matching degree between the media feature of the any media data and the combined behavior features of the various dimensions.
[0142] In one implementation, the generating unit 601 obtains the combined behavior feature of any dimension according to the behavior feature unit vector of any dimension and the shared behavior feature unit vector, including:
[0143] Processing the identity features of the user through the gated network in the trained user representation model to obtain a first weight of the behavior feature unit vector of any dimension and a second weight of the shared behavior feature unit vector, wherein the first weight is used to indicate the contribution degree of the user to any dimension;
[0144] A combined behavior feature of any dimension is obtained according to the behavior feature unit vector of any dimension, the shared behavior feature unit vector, the first weight, and the second weight.
[0145] In one implementation, the comparison unit 602 is further configured to include:
[0146] Performing dimensionality reduction processing on the behavior feature unit vector of any dimension through the sub-tower corresponding to any dimension in the trained user representation model to obtain the behavior feature unit vector after dimensionality reduction in any dimension;
[0147] For any media data in the media database, comparing the behavior feature unit vectors of each dimension with the media features of the any media data to obtain the matching degree between the media features of the any media data and the behavior feature unit vectors of each dimension, includes:
[0148] For any media data in the media database, the behavior feature unit vectors after dimension reduction of each dimension are respectively compared with the media features of the any media data to obtain the matching degree between the media features of the any media data and the behavior feature unit vectors after dimension reduction of each dimension;
[0149] The obtaining the matching degree between the user and the any media data according to the matching degree between the media feature of the any media data and the behavior feature unit vectors of each dimension includes:
[0150] The matching degree between the user and the any media data is obtained according to the matching degree between the media features of the any media data and the behavior feature unit vectors after dimension reduction in each dimension.
[0151] In one implementation, the comparison unit 602 is further configured to include:
[0152] splicing the behavior feature unit vector of any dimension and the behavior feature unit vector of any dimension after dimensionality reduction to obtain a spliced behavior feature unit vector of any dimension;
[0153] For any media data in the media database, comparing the reduced-dimensional behavior feature unit vectors of each dimension with the media features of the any media data, respectively, to obtain a matching degree between the media features of the any media data and the reduced-dimensional behavior feature unit vectors of each dimension, includes:
[0154] For any media data in the media database, the spliced behavior feature unit vectors of each dimension are respectively compared with the media features of the any media data to obtain a matching degree between the media features of the any media data and the spliced behavior feature unit vectors of each dimension;
[0155] The obtaining the matching degree between the user and the any media data according to the matching degree between the media feature of the any media data and the behavior feature unit vector after dimension reduction of each dimension includes:
[0156] The matching degree between the user and the any media data is obtained according to the matching degree between the media features of the any media data and the spliced behavior feature unit vectors of the various dimensions.
[0157] In one implementation, the generating unit 601 obtains the matching degree between the user and the any media data according to the matching degree between the media feature of the any media data and the behavior feature unit vectors of each dimension, including:
[0158] According to the weight of each dimension, a weighted sum is performed on the matching degree between the behavior feature unit vector of each dimension and the media feature of any media data to obtain the matching degree between the user and the any media data.
[0159] In one implementation, the generating unit 601 is further configured to include:
[0160] Generate training behavior feature unit vectors of each dimension based on the behavior features of the training user through an expert network corresponding to each dimension of the multiple dimensions included in the user representation model; the training behavior features are obtained based on the interactive behavior data generated during the interaction between the training user and the training media data;
[0161] For any training media data in the training media database, the training behavior feature unit vectors of the respective dimensions are respectively compared with the media features of the any training media data to obtain a matching degree between the media features of the any training media data and the training behavior feature unit vectors of the respective dimensions;
[0162] Obtaining a matching degree between the training user and the any training media data according to a matching degree between a media feature of the any training media data and a training behavior feature unit vector of each dimension;
[0163] The user representation model is trained according to the matching degree between the training user and any training media data and the preset matching degree to obtain the trained user representation model; the expert network corresponding to each dimension of the multiple dimensions included in the trained user representation model is used to generate the behavior feature unit vector of each dimension based on the user's behavior features.
[0164] In the embodiment of the present application, the generation unit 601 generates the behavior feature unit vector of each dimension based on the behavior features of the user through the expert network corresponding to each dimension in the multiple dimensions included in the trained user representation model; the comparison unit 602 compares the behavior feature unit vector of each dimension with the media features of any media data in the media database, and obtains the matching degree between the media features of any media data and the behavior feature unit vector of each dimension; the generation unit 601 obtains the matching degree between the user and any media data according to the matching degree between the media features of any media data and the behavior feature unit vector of each dimension; the identification unit 603 identifies the target media data with a matching degree with the user higher than a preset threshold as the user's recalled media data. Therefore, generating the behavior feature unit vector of each dimension based on the user's behavior features can accurately represent the user's interest points; calculating the matching degree with the media data according to the behavior feature unit vector of each dimension, and finally determining the recalled media data, can improve the push accuracy of the media data.
[0165] See also Figure 7 , Figure 7 7 is a schematic diagram of the structure of a computer device provided in an embodiment of the present application. The computer device in the embodiment of the present application includes a power supply module and other structures, and includes a processor 701, a memory 702, and a communication interface 703. The processor 701, the memory 702, and the communication interface 703 can exchange data, and the processor 701 implements the corresponding media data recall method.
[0166] The memory 702 may include a volatile memory, such as a random-access memory (RAM); the memory 702 may also include a non-volatile memory, such as a flash memory, a solid-state drive (SSD), etc.; the memory 702 may also include a combination of the above-mentioned types of memory.
[0167] The processor 701 may be a central processing unit (CPU). The processor 701 may also be a combination of a CPU and a GPU. In a computer device, multiple CPUs and GPUs may be included as needed to recall corresponding media data. In one embodiment, the memory 702 is used to store program instructions. The processor 701 may call program instructions to implement various methods involved in the embodiments of the present application.
[0168] In a first possible implementation, the processor 701 of the computer device calls the program instructions stored in the memory 702 to generate a behavior feature unit vector of each dimension based on the user's behavior features through an expert network corresponding to each dimension of the multiple dimensions included in the trained user representation model; the behavior features are obtained based on the interactive behavior data generated during the interaction between the user and the media data; for any media data in the media database, the behavior feature unit vector of each dimension is compared with the media features of any media data respectively to obtain the matching degree between the media features of any media data and the behavior feature unit vector of each dimension; based on the matching degree between the media features of any media data and the behavior feature unit vector of each dimension, the matching degree between the user and any media data is obtained; and the target media data whose matching degree with the user is higher than a preset threshold is identified as the user's recalled media data.
[0169] In one implementation, the processor 701 may further perform the following operations:
[0170] Generate a shared behavior feature unit vector based on the behavior feature through a shared expert network in the trained user representation model;
[0171] Obtaining a combined behavior feature of any dimension according to the behavior feature unit vector of any dimension and the shared behavior feature unit vector;
[0172] For any media data in the media database, comparing the behavior feature unit vectors of each dimension with the media features of the any media data to obtain the matching degree between the media features of the any media data and the behavior feature unit vectors of each dimension, includes:
[0173] For any media data in the media database, respectively compare the combined behavior features of each dimension with the media features of the any media data to obtain a matching degree between the media features of the any media data and the combined behavior features of each dimension;
[0174] The obtaining the matching degree between the user and the any media data according to the matching degree between the media feature of the any media data and the behavior feature unit vectors of each dimension includes:
[0175] The matching degree between the user and the any media data is obtained according to the matching degree between the media feature of the any media data and the combined behavior features of the various dimensions.
[0176] In one implementation, the processor 701 obtains the combined behavior feature of any dimension according to the behavior feature unit vector of any dimension and the shared behavior feature unit vector, and may perform the following operations:
[0177] Processing the identity features of the user through the gated network in the trained user representation model to obtain a first weight of the behavior feature unit vector of any dimension and a second weight of the shared behavior feature unit vector, wherein the first weight is used to indicate the contribution degree of the user to any dimension;
[0178] A combined behavior feature of any dimension is obtained according to the behavior feature unit vector of any dimension, the shared behavior feature unit vector, the first weight, and the second weight.
[0179] In one implementation, the processor 701 may further perform the following operations:
[0180] Performing dimensionality reduction processing on the behavior feature unit vector of any dimension through the sub-tower corresponding to any dimension in the trained user representation model to obtain the behavior feature unit vector after dimensionality reduction in any dimension;
[0181] For any media data in the media database, comparing the behavior feature unit vectors of each dimension with the media features of the any media data to obtain the matching degree between the media features of the any media data and the behavior feature unit vectors of each dimension, includes:
[0182] For any media data in the media database, the behavior feature unit vectors after dimension reduction of each dimension are respectively compared with the media features of the any media data to obtain the matching degree between the media features of the any media data and the behavior feature unit vectors after dimension reduction of each dimension;
[0183] The obtaining the matching degree between the user and the any media data according to the matching degree between the media feature of the any media data and the behavior feature unit vectors of each dimension includes:
[0184] The matching degree between the user and the any media data is obtained according to the matching degree between the media features of the any media data and the behavior feature unit vectors after dimension reduction in each dimension.
[0185] In one implementation, the processor 701 may further perform the following operations:
[0186] splicing the behavior feature unit vector of any dimension and the behavior feature unit vector of any dimension after dimensionality reduction to obtain a spliced behavior feature unit vector of any dimension;
[0187] For any media data in the media database, comparing the reduced-dimensional behavior feature unit vectors of each dimension with the media features of the any media data, respectively, to obtain a matching degree between the media features of the any media data and the reduced-dimensional behavior feature unit vectors of each dimension, includes:
[0188] For any media data in the media database, the spliced behavior feature unit vectors of each dimension are respectively compared with the media features of the any media data to obtain a matching degree between the media features of the any media data and the spliced behavior feature unit vectors of each dimension;
[0189] The obtaining the matching degree between the user and the any media data according to the matching degree between the media feature of the any media data and the behavior feature unit vector after dimension reduction of each dimension includes:
[0190] The matching degree between the user and the any media data is obtained according to the matching degree between the media features of the any media data and the spliced behavior feature unit vectors of the various dimensions.
[0191] In one implementation, the processor 701 obtains the matching degree between the user and the any media data according to the matching degree between the media feature of the any media data and the behavior feature unit vectors of each dimension, and may perform the following operations:
[0192] According to the weight of each dimension, a weighted sum is performed on the matching degree between the behavior feature unit vector of each dimension and the media feature of any media data to obtain the matching degree between the user and the any media data.
[0193] In one implementation, the processor 701 may further perform the following operations:
[0194] Generate training behavior feature unit vectors of each dimension based on the behavior features of the training user through an expert network corresponding to each dimension of the multiple dimensions included in the user representation model; the training behavior features are obtained based on the interactive behavior data generated during the interaction between the training user and the training media data;
[0195] For any training media data in the training media database, the training behavior feature unit vectors of the respective dimensions are respectively compared with the media features of the any training media data to obtain a matching degree between the media features of the any training media data and the training behavior feature unit vectors of the respective dimensions;
[0196] Obtaining a matching degree between the training user and the any training media data according to a matching degree between a media feature of the any training media data and a training behavior feature unit vector of each dimension;
[0197] The user representation model is trained according to the matching degree between the training user and any training media data and the preset matching degree to obtain the trained user representation model; the expert network corresponding to each dimension of the multiple dimensions included in the trained user representation model is used to generate the behavior feature unit vector of each dimension based on the user's behavior features.
[0198] In the embodiment of the present application, the processor 701 generates a behavior feature unit vector of each dimension based on the user's behavior features through an expert network corresponding to each dimension in the multiple dimensions included in the trained user representation model; for any media data in the media database, the behavior feature unit vector of each dimension is respectively compared with the media features of any media data to obtain the matching degree between the media features of any media data and the behavior feature unit vector of each dimension; according to the matching degree between the media features of any media data and the behavior feature unit vector of each dimension, the matching degree between the user and any media data is obtained; the target media data whose matching degree with the user is higher than a preset threshold is identified as the user's recalled media data. Therefore, the user's interest points can be accurately represented by generating the behavior feature unit vector of each dimension through the user's behavior features; the matching degree with the media data is calculated according to the behavior feature unit vector of each dimension, and the recalled media data is finally determined, which can improve the push accuracy of the media data.
[0199] Those skilled in the art can understand that to implement all or part of the processes in the above-mentioned embodiments, the processes can be completed by computer programs to instruct related hardware, and the programs can be stored in computer-readable storage media. When the programs are executed, they can include the processes of the above-mentioned method embodiments. The aforementioned storage media include: ROM or random access memory RAM, magnetic disk or optical disk and other media that can store program codes.
[0200] The above disclosure is only part of the embodiments of the present application, which certainly cannot be used to limit the scope of rights of the present application. Ordinary technicians in this field can understand that all or part of the processes of implementing the above embodiments and equivalent changes made according to the claims of the present application are still within the scope covered by the present application.
Claims
1. A method for recalling media data, characterized in that: include: Generate a behavior feature unit vector of each dimension based on the behavior feature of the user through an expert network corresponding to each dimension of the multiple dimensions included in the trained user representation model; the behavior feature is obtained based on the interactive behavior data generated during the interaction between the user and the media data; For any media data in the media database, the behavior feature unit vectors of each dimension are respectively compared with the media features of the any media data to obtain the matching degree between the media features of the any media data and the behavior feature unit vectors of each dimension; Obtaining a matching degree between the user and the any media data according to a matching degree between a media feature of the any media data and a behavior feature unit vector of each dimension; The target media data whose matching degree with the user is higher than a preset threshold is identified as the recalled media data of the user.
2. The method according to claim 1, characterized in that The method further comprises: Generate a shared behavior feature unit vector based on the behavior feature through a shared expert network in the trained user representation model; Obtaining a combined behavior feature of any dimension according to the behavior feature unit vector of any dimension and the shared behavior feature unit vector; For any media data in the media database, comparing the behavior feature unit vectors of each dimension with the media features of the any media data to obtain the matching degree between the media features of the any media data and the behavior feature unit vectors of each dimension, includes: For any media data in the media database, respectively compare the combined behavior features of each dimension with the media features of the any media data to obtain a matching degree between the media features of the any media data and the combined behavior features of each dimension; The obtaining the matching degree between the user and the any media data according to the matching degree between the media feature of the any media data and the behavior feature unit vectors of each dimension includes: The matching degree between the user and the any media data is obtained according to the matching degree between the media feature of the any media data and the combined behavior features of the various dimensions.
3. The method according to claim 2, characterized in that The obtaining of the combined behavior feature of any dimension according to the behavior feature unit vector of any dimension and the shared behavior feature unit vector includes: Processing the identity features of the user through the gated network in the trained user representation model to obtain a first weight of the behavior feature unit vector of any dimension and a second weight of the shared behavior feature unit vector, wherein the first weight is used to indicate the contribution degree of the user to any dimension; A combined behavior feature of any dimension is obtained according to the behavior feature unit vector of any dimension, the shared behavior feature unit vector, the first weight, and the second weight.
4. The method according to claim 1, characterized in that The method further comprises: Performing dimensionality reduction processing on the behavior feature unit vector of any dimension through the sub-tower corresponding to any dimension in the trained user representation model to obtain the behavior feature unit vector after dimensionality reduction in any dimension; For any media data in the media database, comparing the behavior feature unit vectors of each dimension with the media features of the any media data to obtain the matching degree between the media features of the any media data and the behavior feature unit vectors of each dimension, includes: For any media data in the media database, the behavior feature unit vectors after dimension reduction of each dimension are respectively compared with the media features of the any media data to obtain the matching degree between the media features of the any media data and the behavior feature unit vectors after dimension reduction of each dimension; The obtaining the matching degree between the user and the any media data according to the matching degree between the media feature of the any media data and the behavior feature unit vectors of each dimension includes: The matching degree between the user and the any media data is obtained according to the matching degree between the media features of the any media data and the behavior feature unit vectors after dimension reduction in each dimension.
5. The method according to claim 4, characterized in that The method further comprises: splicing the behavior feature unit vector of any dimension and the behavior feature unit vector of any dimension after dimensionality reduction to obtain a spliced behavior feature unit vector of any dimension; For any media data in the media database, comparing the reduced-dimensional behavior feature unit vectors of each dimension with the media features of the any media data, respectively, to obtain a matching degree between the media features of the any media data and the reduced-dimensional behavior feature unit vectors of each dimension, includes: For any media data in the media database, the spliced behavior feature unit vectors of each dimension are respectively compared with the media features of the any media data to obtain a matching degree between the media features of the any media data and the spliced behavior feature unit vectors of each dimension; The obtaining the matching degree between the user and the any media data according to the matching degree between the media feature of the any media data and the behavior feature unit vector after dimension reduction of each dimension includes: The matching degree between the user and the any media data is obtained according to the matching degree between the media features of the any media data and the spliced behavior feature unit vectors of the various dimensions.
6. The method according to claim 1, characterized in that The obtaining the matching degree between the user and the any media data according to the matching degree between the media feature of the any media data and the behavior feature unit vectors of each dimension includes: According to the weight of each dimension, a weighted sum is performed on the matching degree between the behavior feature unit vector of each dimension and the media feature of any media data to obtain the matching degree between the user and the any media data.
7. The method according to claim 1, characterized in that The method further comprises: Generate training behavior feature unit vectors of each dimension based on the behavior features of the training user through an expert network corresponding to each dimension of the multiple dimensions included in the user representation model; the training behavior features are obtained based on the interactive behavior data generated during the interaction between the training user and the training media data; For any training media data in the training media database, the training behavior feature unit vectors of the respective dimensions are respectively compared with the media features of the any training media data to obtain a matching degree between the media features of the any training media data and the training behavior feature unit vectors of the respective dimensions; Obtaining a matching degree between the training user and the any training media data according to a matching degree between a media feature of the any training media data and a training behavior feature unit vector of each dimension; The user representation model is trained according to the matching degree between the training user and any training media data and the preset matching degree to obtain the trained user representation model; the expert network corresponding to each dimension of the multiple dimensions included in the trained user representation model is used to generate the behavior feature unit vector of each dimension based on the user's behavior features.
8. A media data recall device, characterized in that: include: A generating unit, configured to generate a behavior feature unit vector of each dimension based on the behavior feature of the user through an expert network corresponding to each dimension of the multiple dimensions included in the trained user representation model; the behavior feature is obtained based on the interactive behavior data generated during the interaction between the user and the media data; A comparison unit, for comparing the behavior feature unit vectors of each dimension with the media features of any media data in the media database, to obtain a matching degree between the media features of any media data and the behavior feature unit vectors of each dimension; The generating unit is further configured to obtain a matching degree between the user and the any media data according to a matching degree between a media feature of the any media data and a behavior feature unit vector of each dimension; The identification unit is used to identify the target media data whose matching degree with the user is higher than a preset threshold as the recalled media data of the user.
9. A computer device, characterized in that: The computer device includes a memory, a communication interface and a processor, wherein the memory, the communication interface and the processor are interconnected; the memory stores a computer program, and the processor calls the computer program stored in the memory to implement the media data recall method described in any one of claims 1 to 7.
10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the method for recalling media data according to any one of claims 1 to 7 is implemented.
11. A computer program product, characterized in that The computer program product includes a computer program, which is stored in a computer storage medium; a processor of a computer device reads the computer program from the computer storage medium, and the processor executes the computer program, so that the computer device executes the media data recall method as described in any one of claims 1 to 7.
Citation Information
Cited By
Object recommendation method, device and system, computing equipment and storage medium
CN119917749A
