Media asset recommendation method and device, equipment, storage medium and program product

Abstract

The invention provides a media asset recommendation method and device, equipment, a storage medium and a program product, and is applied to the technical field of computers. Obtaining user information of a target user, historical playing media asset information of the target user and associated information of a media asset recommendation request of the target user; obtaining media asset information of media assets to be recommended; inputting the user information of the target user, the historical playing media asset information of the target user, the associated information of the media asset recommendation request of the target user and the media asset information of the media assets to be recommended into a media asset recommendation model, and obtaining a predicted click rate of the media assets to be recommended, which is output by the media asset recommendation model, of the target user; and based on the predicted click rate of the target user on the to-be-recommended media assets, determining target media assets recommended to the target user from the to-be-recommended media assets, and recommending the target media assets. According to the method, the accuracy of recommending the media assets to the user can be improved.

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular to a media asset recommendation method, apparatus, device, storage medium, and program product. Background Technology

[0002] With the development of big data and artificial intelligence technologies, many internet applications have incorporated recommendation systems as a core component. These systems can push personalized media content based on user behavior and preferences. Most applications use a "feed" approach to content delivery, meaning users can continuously scroll down the page to receive new media recommendations. This model is commonly used in mobile applications and websites because it provides a seamless and seemingly endless content browsing experience.

[0003] Existing media asset recommendation methods, in order to meet users' novelty needs, directly filter out media assets that users have previously viewed or clicked, or demote repeatedly exposed media assets. However, different users have different needs for novelty; some users may prefer familiar content, while others like to constantly seek new experiences. Therefore, simply filtering or demoting cannot accurately meet the preferences of all users. Consequently, existing media asset recommendation methods have low accuracy in pushing media assets to users. Summary of the Invention

[0004] This application provides a media asset recommendation method and apparatus to solve the problem of low positioning accuracy in existing media asset recommendation methods.

[0005] To solve the above-mentioned technical problems, this application is implemented as follows:

[0006] In a first aspect, embodiments of this application provide a media asset recommendation method, which includes:

[0007] Upon receiving a media asset recommendation request from a target user, the system obtains the target user's user information, the target user's historical media asset information, and the associated information of the target user's media asset recommendation request.

[0008] Obtain media asset information for media assets to be recommended;

[0009] The target user's user information, the target user's historical playback media asset information, the target user's media asset recommendation request association information, and the media asset information of the media asset to be recommended are input into the media asset recommendation model. The predicted click-through rate of the target user for the media asset to be recommended is obtained from the output of the media asset recommendation model. The media asset recommendation model is obtained by training a neural network model using the user information of multiple training sample users, the historical playback media asset information of the multiple training sample users, the association information of the media asset recommendation requests of the multiple training sample users, the media asset information of multiple training sample media, and the actual click situation of each training sample user for each training sample media.

[0010] Based on the predicted click-through rate of the target user for the media asset to be recommended, a target media asset is determined from the media assets to be recommended and recommended to the target user.

[0011] Optionally, the media asset recommendation model includes a user novelty intent recognition module, a user multi-interest extraction module, and a media asset novelty representation module;

[0012] The step of inputting the target user's user information, the target user's historical media asset information, the association information of the target user's media asset recommendation request, and the media asset information of the media asset to be recommended into the media asset recommendation model, and obtaining the predicted click-through rate of the target user for the media asset to be recommended output by the media asset recommendation model, includes:

[0013] The user information of the target user and the association information of the target user's media asset recommendation request are input into the user novelty intent recognition module to obtain the novelty intent representation of the target user;

[0014] The target user's historical media asset information is input into the user's multi-interest extraction module to obtain the target user's multi-interest representation;

[0015] The media asset information of the media asset to be recommended is input into the media asset novelty representation module to obtain the novelty representation of the media asset to be recommended by the target user;

[0016] Based on the target user's novelty intent representation, the target user's multi-interest representation, and the target user's novelty representation of the media asset to be recommended, the predicted click-through rate of the target user for the media asset to be recommended is determined.

[0017] Optionally, the step of inputting the user information of the target user and the association information of the target user's media asset recommendation request into the user novelty intent recognition module to obtain the novelty intent representation of the target user includes:

[0018] The user information of the target user is input into the first embedding layer of the novelty intent recognition module, and the user feature representation of the target user is obtained by using the multilayer perceptron (MLP) network of the novelty intent recognition module.

[0019] The association information of the target user's media asset recommendation request is input into the second embedding layer of the novelty intent recognition module, and the MLP network of the novelty intent recognition module is used to obtain the target user's recommendation request behavior representation;

[0020] The user feature representation and the recommendation request behavior representation are input into the MLP network of the novelty intent recognition module, and the novelty intent representation of the target user is calculated using the first activation function.

[0021] Optionally, after inputting the user information of the target user and the association information of the target user's media asset recommendation request into the user novelty intent recognition module to obtain the novelty intent representation of the target user, the method further includes:

[0022] Based on the novelty intent representation of the target user, the non-novelty intent score of the target user is calculated using the second activation function;

[0023] The historical playback media asset information includes media asset information of multiple media assets played by the target user within a preset historical time period; the step of inputting the target user's historical playback media asset information into the user multi-interest extraction module to obtain the target user's multi-interest representation includes:

[0024] The media asset information of multiple media assets played by the target user within a preset historical time period is input into the embedding layer of the user multi-interest extraction module to obtain the media asset information representation of the multiple media assets.

[0025] The media asset information of the multiple media assets is concatenated in chronological order to obtain the historical playback behavior sequence of the target user;

[0026] The historical playback behavior sequence of the target user is input into the attention layer of the user multi-interest extraction module to obtain the initial multi-interest representation of the target user;

[0027] The initial multi-interest representation is weighted using the non-novelty intent score to obtain the multi-interest representation of the target user.

[0028] Optionally, after inputting the user information of the target user and the association information of the target user's media asset recommendation request into the user novelty intent recognition module to obtain the novelty intent representation of the target user, the method further includes:

[0029] Based on the novelty intent representation of the target user, the novelty intent score of the target user is calculated using the second activation function;

[0030] The media asset information of the media asset to be recommended is input into the media asset novelty representation module to obtain the novelty representation of the media asset to be recommended by the target user, including:

[0031] The media asset information of the media asset to be recommended is input into the embedding layer of the media asset novelty representation module to obtain the first media asset feature representation of the media asset to be recommended;

[0032] The novelty status marker of the media asset to be recommended and the first media asset feature representation of the media asset to be recommended are concatenated to obtain the second media asset feature representation of the media asset to be recommended. The novelty status marker of the media asset to be recommended is obtained based on the target user's operation behavior on the media asset to be recommended in different historical periods.

[0033] The second media asset feature representation is input into the compression and incentive network SENet model of the media asset novelty representation module. After compression and incentive, the weight vector of the media asset information of the media asset to be recommended is obtained.

[0034] The second media asset feature representation is weighted using the weight vector to obtain the third media asset feature representation of the media asset to be recommended;

[0035] The novelty intent score is used to weight the feature representation of the third media asset to obtain the novelty representation of the media asset to be recommended by the target user.

[0036] Optionally, determining the predicted click-through rate of the target user for the media asset to be recommended based on the target user's novelty intent representation, the target user's multi-interest representation, and the target user's novelty representation of the media asset to be recommended includes:

[0037] The multi-interest representation of the target user and the novelty representation of the target user for the media asset to be recommended are concatenated and then input into the MLP network of the media asset recommendation model to obtain the fused information representation;

[0038] The novelty representation of the target user to the media asset to be recommended is input into the MLP network of the media asset recommendation model, and the novelty score of the target user to the media asset to be recommended is calculated using the third activation function.

[0039] The novelty score of the target user for the media asset to be recommended is used to calibrate the fused information representation, and the calibrated fused information representation is input into the MLP network of the media asset recommendation model. The predicted click-through rate of the target user for the media asset to be recommended is calculated using the third activation function.

[0040] Optionally, the media asset recommendation model is trained in the following manner:

[0041] The user information of the training sample users, the historical playback media asset information of the training sample users, the association information of the media asset recommendation requests of the training sample users, and the media asset information of the training sample media are input into the media asset recommendation model to obtain the predicted click-through rate of the training sample users for the training sample media.

[0042] A binary cross-entropy loss function is constructed using the predicted click-through rate of the training sample users on the training sample media assets and the actual click-through rate of the training sample users on the training sample media assets.

[0043] The media asset recommendation model is trained with the goal of minimizing the binary cross-entropy loss function.

[0044] Secondly, embodiments of this application also provide a media asset recommendation device, which includes:

[0045] The first acquisition module is used to acquire, upon receiving a media asset recommendation request from a target user, the target user's user information, the target user's historical playback media asset information, and the associated information of the target user's media asset recommendation request.

[0046] The second acquisition module is used to acquire media asset information of the media assets to be recommended;

[0047] The third acquisition module is used to input the user information of the target user, the historical playback media asset information of the target user, the association information of the media asset recommendation request of the target user, and the media asset information of the media asset to be recommended into the media asset recommendation model, and to obtain the predicted click-through rate of the target user for the media asset to be recommended output by the media asset recommendation model. The media asset recommendation model is obtained by training a neural network model using the user information of multiple training sample users, the historical playback media asset information of the multiple training sample users, the association information of the media asset recommendation request of the multiple training sample users, the media asset information of multiple training sample media, and the actual click situation of each training sample user for each training sample media.

[0048] The first recommendation module is used to determine the target media asset to be recommended to the target user from the media assets to be recommended based on the predicted click-through rate of the target user on the media assets to be recommended, and to recommend the target media asset.

[0049] Thirdly, embodiments of this application also provide an electronic device, which includes a processor, the processor being used for:

[0050] Upon receiving a media asset recommendation request from a target user, the system obtains the target user's user information, the target user's historical media asset information, and the associated information of the target user's media asset recommendation request.

[0051] Obtain media asset information for media assets to be recommended;

[0052] The target user's user information, the target user's historical playback media asset information, the target user's media asset recommendation request association information, and the media asset information of the media asset to be recommended are input into the media asset recommendation model. The predicted click-through rate of the target user for the media asset to be recommended is obtained from the output of the media asset recommendation model. The media asset recommendation model is obtained by training a neural network model using the user information of multiple training sample users, the historical playback media asset information of the multiple training sample users, the association information of the media asset recommendation requests of the multiple training sample users, the media asset information of multiple training sample media, and the actual click situation of each training sample user for each training sample media.

[0053] Based on the predicted click-through rate of the target user for the media asset to be recommended, a target media asset is determined from the media assets to be recommended and recommended to the target user.

[0054] Fourthly, embodiments of this application also provide a media asset recommendation device, including a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the steps of the media asset recommendation method described above.

[0055] Fifthly, embodiments of this application also provide a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the media asset recommendation method described above.

[0056] A sixth aspect provides a computer program product including computer instructions that, when executed by a processor, implement the steps of the media asset recommendation method as described in the first aspect.

[0057] The media asset recommendation method of this application embodiment includes, upon receiving a media asset recommendation request from a target user, obtaining the target user's user information, the target user's historical playback media asset information, and the association information of the target user's media asset recommendation request; obtaining the media asset information of the media asset to be recommended; inputting the target user's user information, the target user's historical playback media asset information, the association information of the target user's media asset recommendation request, and the media asset information of the media asset to be recommended into a media asset recommendation model, and obtaining the predicted click-through rate of the target user for the media asset to be recommended output by the media asset recommendation model. The media asset recommendation model is obtained by training a neural network model using the user information of multiple training sample users, the historical playback media asset information of multiple training sample users, the association information of the media asset recommendation requests of multiple training sample users, the media asset information of multiple training sample media, and the actual click situation of each training sample user for each training sample media; and determining the target media asset to be recommended to the target user from the media asset to be recommended based on the predicted click-through rate of the target user for the media asset to be recommended, and recommending the target media asset. This method can identify users' demand for media asset novelty based on user information, users' historical media asset playback information, and the correlation information of users' media asset recommendation requests. Using this information to determine the click-through rate of media assets can improve the accuracy of media asset recommendations to users. Attached Figure Description

[0058] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments of this application will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0059] Figure 1 This is a flowchart of the media asset recommendation method provided in the embodiments of this application;

[0060] Figure 2 This is a model structure diagram of the media asset recommendation method provided in the embodiments of this application;

[0061] Figure 3 This is a structural diagram of a media resource recommendation device provided in an embodiment of this application;

[0062] Figure 4 This is a structural diagram of an electronic device provided in an embodiment of this application. Detailed Implementation

[0063] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0064] This application provides a media asset recommendation method. See also... Figure 1 , Figure 1 This is a flowchart of the media asset recommendation method provided in the embodiments of this application, such as... Figure 1 As shown, it includes the following steps:

[0065] Step 101: Upon receiving a media asset recommendation request from a target user, obtain the target user's user information, the target user's historical playback media asset information, and the association information of the target user's media asset recommendation request;

[0066] In this step, the target user's information includes the target user's age segmentation information, the target user's channel preferences (e.g., movies or TV series), the target user's tag preferences (e.g., romance dramas or suspense dramas), and the target user's type preferences (e.g., urban romance or campus romance). The target user's historical media asset information includes the media asset information of multiple media assets that the target user has played within a preset historical period (e.g., the last 30 days). The associated information of the target user's media asset recommendation request includes the timestamp of the target user's media asset recommendation request, the province where the target user is located, which refresh the target user entered the recommendation screen, the target user's dwell time on the recommendation screen, and the media asset information of the previous media asset clicked by the target user.

[0067] The information above is relevant to identifying users' novelty intent. For example, different age groups typically have different needs for novel experiences. Young people may be more inclined to explore new trends, while older people may prefer classic content or content relevant to their current stage of life. A user's long-term focus on a specific channel can reflect their preferred themes or styles, but occasionally browsing other less frequently visited channels may indicate a desire for novel content outside their usual interests. Specific times or locations may trigger a special need for new content (such as holidays, special events, etc.). Adjusting the freshness of recommendations based on time and location changes can better meet users' needs for new information. Frequent refreshing may indicate that users haven't found the new content they're looking for, while longer dwell times usually indicate satisfaction with the current recommendations. If short browsing is accompanied by frequent refreshes, it may indicate that users crave bolder recommendations.

[0068] Step 102: Obtain the media asset information of the media assets to be recommended;

[0069] In this step, the media asset information to be recommended includes media asset identifier (ID), media asset duration, media asset language, media asset region, number of clicks in the past 1 / 7 / 15 / 30 days, number of plays in the past 1 / 7 / 15 / 30 days, and number of exposures in the past 1 / 7 / 15 / 30 days.

[0070] Step 103: Input the user information of the target user, the historical playback media asset information of the target user, the association information of the media asset recommendation request of the target user, and the media asset information of the media asset to be recommended into the media asset recommendation model, and obtain the predicted click-through rate of the target user for the media asset to be recommended output by the media asset recommendation model. The media asset recommendation model is obtained by training a neural network model using the user information of multiple training sample users, the historical playback media asset information of the multiple training sample users, the association information of the media asset recommendation request of the multiple training sample users, the media asset information of multiple training sample media, and the actual click situation of each training sample user for each training sample media.

[0071] In this step, the media recommendation model is first trained using user information from multiple training sample users, historical playback media asset information from multiple training sample users, association information of media asset recommendation requests from multiple training sample users, media asset information from multiple training sample media, and the actual click data of each training sample user for each training sample media. Specifically:

[0072] The user information of the training sample users, the historical playback media asset information of the training sample users, the association information of the media asset recommendation requests of the training sample users, and the media asset information of the training sample media are input into the media asset recommendation model to obtain the predicted click-through rate of the training sample users for the training sample media.

[0073] A binary cross-entropy loss function is constructed using the predicted click-through rate of training sample users on the training sample media assets and the actual click-through rate of training sample users on the training sample media assets.

[0074] The media asset recommendation model is trained with the goal of minimizing the binary cross-entropy loss function.

[0075] Then, the target user's user information, the target user's historical media asset information, the target user's media asset recommendation request association information, and the media asset information of the media asset to be recommended are input into the trained media asset recommendation model to obtain the target user's predicted click-through rate of the media asset to be recommended.

[0076] Step 104: Based on the predicted click-through rate of the target user for the media asset to be recommended, determine the target media asset to be recommended to the target user from the media assets to be recommended, and recommend the target media asset.

[0077] In this step, a threshold can be preset. If the predicted click-through rate of the target user for the recommended media asset exceeds this threshold, then the media asset will be recommended to the target user.

[0078] Alternatively, if there are multiple media assets to be recommended, the media asset with the highest predicted click-through rate will be selected as the one recommended to the target user.

[0079] In one implementation, upon receiving a media asset recommendation request from a target user, the system acquires the target user's user information, the target user's historical playback media asset information, and the association information of the target user's media asset recommendation request; it also acquires the media asset information of the media asset to be recommended; inputs the target user's user information, the target user's historical playback media asset information, the association information of the target user's media asset recommendation request, and the media asset information of the media asset to be recommended into a media asset recommendation model; and obtains the predicted click-through rate of the target user for the media asset to be recommended, output by the media asset recommendation model. The media asset recommendation model is trained using user information of multiple training sample users, historical playback media asset information of multiple training sample users, association information of media asset recommendation requests of multiple training sample users, media asset information of multiple training sample media, and the actual click situation of each training sample user for each training sample media. Based on the predicted click-through rate of the target user for the media asset to be recommended, the system determines the target media asset to be recommended to the target user from the media asset to be recommended and recommends the target media asset.

[0080] In this implementation, the user's demand for media asset novelty can be identified based on the user's user information, the user's historical media asset playback information, and the association information of the user's media asset recommendation request. Using this information to determine the user's click-through rate for media assets can improve the accuracy of recommending media assets to the user.

[0081] Optionally, the media asset recommendation model includes a user novelty intent recognition module, a user multi-interest extraction module, and a media asset novelty representation module;

[0082] The step of inputting the target user's user information, the target user's historical media asset information, the association information of the target user's media asset recommendation request, and the media asset information of the media asset to be recommended into the media asset recommendation model, and obtaining the predicted click-through rate of the target user for the media asset to be recommended output by the media asset recommendation model, includes:

[0083] The user information of the target user and the association information of the target user's media asset recommendation request are input into the user novelty intent recognition module to obtain the novelty intent representation of the target user;

[0084] The target user's historical media asset information is input into the user's multi-interest extraction module to obtain the target user's multi-interest representation;

[0085] The media asset information of the media asset to be recommended is input into the media asset novelty representation module to obtain the novelty representation of the media asset to be recommended by the target user;

[0086] Based on the target user's novelty intent representation, the target user's multi-interest representation, and the target user's novelty representation of the media asset to be recommended, the predicted click-through rate of the target user for the media asset to be recommended is determined.

[0087] In one implementation, see Figure 2 The media asset recommendation model includes a user novelty intent recognition module, a user multi-interest extraction module, and a media asset novelty representation module. Specifically:

[0088] The user information of the target user and the associated information of the target user's media asset recommendation request are respectively input into the user novelty intent recognition module. The novelty intent representation of the target user is obtained by using the embedding layer, connection layer, flattening layer, multilayer perceptron (MLP), and activation functions such as the softmax function and the sigmoid function in the user novelty intent recognition module.

[0089] The user information of the target user and the associated information of the target user's media asset recommendation request are input into the user novelty intent recognition module. The novelty intent representation of the target user is obtained by using the embedding layer, multi-bead self-attention layer, residual connection and layer normalization (Add&Norm), feed-forward network, etc. of the user novelty intent recognition module.

[0090] The media asset information to be recommended is input into the media asset novelty representation module, and the novelty representation of the media asset to be recommended by the target user is obtained by using the Squeeze-and-Excitation Networks (SENet) model, MLP and Flatten layer;

[0091] Then, based on the obtained novelty intent representation of the target user, multi-interest representation of the target user, and novelty representation of the target user towards the recommended media, the predicted click-through rate (CTR) of the target user towards the recommended media is determined using a concatenate layer, an MLP network, and a sigmoid function.

[0092] In this implementation, through multi-level information processing and feature fusion, it is possible to more accurately predict the click-through rate of target users towards recommended media assets.

[0093] Optionally, the step of inputting the user information of the target user and the association information of the target user's media asset recommendation request into the user novelty intent recognition module to obtain the novelty intent representation of the target user includes:

[0094] The user information of the target user is input into the first embedding layer of the novelty intent recognition module, and the user feature representation of the target user is obtained by using the multilayer perceptron (MLP) network of the novelty intent recognition module.

[0095] The association information of the target user's media asset recommendation request is input into the second embedding layer of the novelty intent recognition module, and the MLP network of the novelty intent recognition module is used to obtain the target user's recommendation request behavior representation;

[0096] The user feature representation and the recommendation request behavior representation are input into the MLP network of the novelty intent recognition module, and the novelty intent representation of the target user is calculated using the first activation function.

[0097] In one implementation, user information X u Information including user age binning, user channel preferences, user tag preferences, and user type preferences is concatenated and input into the first embedding layer of the novelty intent recognition module. This transforms the high-dimensional, sparse, discrete data into a low-dimensional, dense vector representation. Then, an MLP is used to extract deeper features from these vector representations to obtain the user feature representation E. u For details, please refer to the following formula:

[0098] X u =[x u，1 x u，2 , ..., x u，n (1)

[0099] E u =f u (W u X u +b u (2)

[0100] Where n represents the number of user information items, W u and b u f represents the trainable parameters of a feedforward neural network based on an MLP. u Represents the activation function ReLU, E u This represents user characteristics.

[0101] The associated information for the target user's media asset recommendation request includes the timestamp of the request, the target user's province, the number of times the target user entered the recommendation screen, the target user's dwell time on the recommendation screen, and the media asset information of the previous media asset clicked by the target user. All associated information is concatenated and input into the second embedding layer of the novelty intent recognition module, converting the high-dimensional, sparse discrete data into a low-dimensional, dense vector representation. Then, a multilayer perceptron (MLP) is used to extract deeper features from these vector representations to obtain the target user's recommendation request behavior representation E. c For details, please refer to the following formula:

[0102] X c =[x c，1 x c，2 , ..., x c，k (3)

[0103] E c =f c (W c X c +b c (4)

[0104] Where k represents the number of associated information, W c and b c f represents the trainable parameters of a feedforward neural network based on an MLP. c Represents the activation function ReLU, E c This represents the recommendation request behavior of the target user.

[0105] In large-screen recommendation scenarios, as users scroll down the recommendation screen, their novelty intent changes dynamically. Accurately capturing these changes in novelty intent during scrolling is beneficial for accurately recommending media resources to users. Specifically, identifying a user's novelty intent can be achieved by inputting the user's feature representation and recommendation request behavior representation into the MLP network of the novelty intent recognition module, and using a first activation function to calculate the target user's novelty intent representation. See the following calculation formula:

[0106] E e =f1(W1E u +W2E c +b1) (5)

[0107] Among them, E e The novelty intent of the target user is represented by W1, W2 and b1, which represent the trainable parameters of the feedforward neural network based on MLP, and f1 represents the activation function ReLU (i.e., the first activation function).

[0108] In this implementation, both the target user's user information and the associated information of the target user's media asset recommendation request are diverse and complex. Through the embedding layer of the novelty intent recognition module, this information is transformed into a unified low-dimensional dense vector form, thereby eliminating the sparsity in the original data and making subsequent analysis more efficient. MLP possesses powerful nonlinear modeling capabilities, enabling it to capture the interaction relationships between complex features and extract latent patterns, which is crucial for understanding the subtle interactions between different features and behavioral variables. Therefore, this implementation can effectively identify the target user's novelty preference.

[0109] Optionally, after inputting the user information of the target user and the association information of the target user's media asset recommendation request into the user novelty intent recognition module to obtain the novelty intent representation of the target user, the method further includes:

[0110] Based on the novelty intent representation of the target user, the non-novelty intent score of the target user is calculated using the second activation function;

[0111] The historical playback media asset information includes media asset information of multiple media assets played by the target user within a preset historical time period; the step of inputting the target user's historical playback media asset information into the user multi-interest extraction module to obtain the target user's multi-interest representation includes:

[0112] The media asset information of multiple media assets played by the target user within a preset historical time period is input into the embedding layer of the user multi-interest extraction module to obtain the media asset information representation of the multiple media assets.

[0113] The media asset information of the multiple media assets is concatenated in chronological order to obtain the historical playback behavior sequence of the target user;

[0114] The historical playback behavior sequence of the target user is input into the attention layer of the user multi-interest extraction module to obtain the initial multi-interest representation of the target user;

[0115] The initial multi-interest representation is weighted using the non-novelty intent score to obtain the multi-interest representation of the target user.

[0116] In one implementation method, the novelty intent score and non-novelty intent score of the target user can be calculated using a second activation function (softmax function) based on the target user's novelty intent representation, as detailed in the following calculation formula:

[0117] p n p f=Softmax(MLP(E e (6)

[0118] Where, p n p represents the novelty intent score of the target user. f Let p represent the non-novelty intent score of the target user, and p n +p f =1.

[0119] The target user's historical media asset information includes media asset information of multiple media assets played by the target user within a preset historical time period.

[0120] Media asset information includes media asset ID, media asset duration, media asset language, media asset region, number of clicks in the past 1 / 7 / 15 / 30 days, number of plays in the past 1 / 7 / 15 / 30 days, and number of exposures in the past 1 / 7 / 15 / 30 days. First, the media asset ID information is mapped. The media asset ID encrypted using Message-Digest Algorithm 5 (MD5) is mapped to a numeric index ID. Then, the numeric index ID is fed into the embedding layer of the user multi-interest extraction module, converting the index ID into an embedding vector. Finally, the other media asset information is fed into the embedding layer of the user multi-interest extraction module. The result is concatenated with the embedding vector, and the media asset information representation is obtained using an MLP network. The specific process can be seen in the following formula:

[0121] index i =look_up(id) i (7)

[0122] emd i =embedding(index) i (8)

[0123] X i =[x i，1 x i，2 , ..., x i，m (9)

[0124] E i =concat[emd i f i (W i X i +b i (10)

[0125] Among them, id i This represents the media asset ID, look_up() represents the mapping function, and index... iThis represents the index id after the mapping function, where embedding() represents the embedding function, and emd i X represents the embedding vector after the embedding function. i This represents all media asset information except for the media asset ID, where m represents the number of media asset information items, and W represents the total number of items. i and b i f represents the trainable parameters of a feedforward neural network based on an MLP. i Represents the activation function ReLU, concat[] represents the concatenation function, and E i This indicates media asset information.

[0126] All media assets played by the target user within a preset historical period (e.g., within 30 days) are concatenated in chronological order to form a sequence of the target user's historical playback behavior. See the formula below for details:

[0127] X b =[item b，1 ;item b，2 ... item b，t (11)

[0128] Where t represents the length of the target user's playback behavior, with a maximum length of 50. If the maximum length is exceeded, the content will be truncated. b，i This represents the t-th media asset after sorting by user click to play. Each media asset in the sequence is expressed according to formulas (7)(8)(9)(10), and then concatenated together. The specific formulas are shown below:

[0129] E b =concat[E b，1 E b，2 ;...;E b，t (12)

[0130] Among them, E b This represents the historical playback behavior sequence of the target user, concat[] represents the concatenation function, and E b，t This represents the t-th media asset after sorting by the target user's click-to-play behavior.

[0131] Then, the user multi-interest extraction module employs an advanced bidirectional transformer encoder to extract diverse user interests and capture complex relationships within the sequence. Each layer of the transformer encoder contains two modules: multi-head self-attention and a feedforward neural network. By jointly focusing on information from various potential subspaces, it models user preferences from multiple interest views. The specific process is illustrated in the following calculation formula:

[0132] MHSA(F l )=[head1, head2,..., head h W o (13)

[0133] Q = F l W i Q (14)

[0134]

[0135] Among them, F l W represents the input of the l-th layer. o This represents the learnable parameters for each head. This represents the normalization scaling factor. and This represents the mapping matrix.

[0136] Using the multi-head self-attention mechanism in formulas (13)(14)(15)(16)(17) above, the historical playback behavior sequence of the target user, E, is represented. b Extract the target user's multiple interest preferences and represent them in different headers to obtain... Where i represents the i-th head and l represents the l-th layer.

[0137] Then, an MLP network is used to fuse multiple interest representations and merge them, and the layer-normalized residuals are used for connection to achieve smooth and stable gradient updates. The specific process can be seen in the following formula:

[0138]

[0139] in, This represents the output of the l-th layer. and This represents the trainable parameters of a feedforward neural network based on an MLP.

[0140] The aforementioned steps yield an unweighted initial multi-interest representation of the target users. Then, the non-novelty intention score p calculated in formula (6) is used. f The initial multi-interest expressions of the target users are weighted, and the specific process is shown in the following formula:

[0141]

[0142] Among them, E x This represents the target user's multiple interests, which are weighted by the target user's non-novelty intent score.

[0143] In this implementation, information on multiple media assets played by the target user during a historical period is input into the embedding layer of the user multi-interest extraction module and concatenated chronologically. This helps to depict the user's interest trajectory, and this time-series representation facilitates the identification of topics or content types that the target user consistently focuses on. The attention layer of the user multi-interest extraction module, by processing the target user's historical playback behavior sequence, can effectively capture multiple interests that the user may simultaneously possess (e.g., different degrees of preference for different types of content). By calculating the target user's non-novelty intent score, it is possible to assess whether the user, given a situation, prefers stable, known favorite content or is willing to accept new recommendations. The initial multi-interest representation is weighted using the target user's non-novelty intent score, and the weighted multi-interest representation can be used to more accurately match the correlation between the target user and the media assets. Thus, if the target user's non-novelty intent score is high, the recommendation strategy can prioritize content types that have successfully attracted the user's attention in the past, while incorporating current preferences into the prediction, allowing for a more accurate estimation of the target user's click-through rate for recommended media assets.

[0144] Optionally, after inputting the user information of the target user and the association information of the target user's media asset recommendation request into the user novelty intent recognition module to obtain the novelty intent representation of the target user, the method further includes:

[0145] Based on the novelty intent representation of the target user, the novelty intent score of the target user is calculated using the second activation function;

[0146] The step of inputting the media asset information of the media asset to be recommended into the media asset novelty representation module to obtain the novelty representation of the media asset to be recommended by the target user includes:

[0147] The media asset information of the media asset to be recommended is input into the embedding layer of the media asset novelty representation module to obtain the first media asset feature representation of the media asset to be recommended;

[0148] The novelty status marker of the media asset to be recommended and the first media asset feature representation of the media asset to be recommended are concatenated to obtain the second media asset feature representation of the media asset to be recommended. The novelty status marker of the media asset to be recommended is obtained based on the target user's operation behavior on the media asset to be recommended in different historical periods.

[0149] The second media asset feature representation is input into the compression and incentive network SENet model of the media asset novelty representation module. After compression and incentive, the weight vector of the media asset information of the media asset to be recommended is obtained.

[0150] The second media asset feature representation is weighted using the weight vector to obtain the third media asset feature representation of the media asset to be recommended;

[0151] The novelty intent score is used to weight the feature representation of the third media asset to obtain the novelty representation of the media asset to be recommended by the target user.

[0152] In one implementation method, referring to the description of the foregoing embodiments, the novelty intent score of the target user is calculated using formula (6). The media asset information of the media asset to be recommended is input into the embedding layer of the media asset novelty representation module, and the first media asset feature representation of the media asset to be recommended is obtained using formulas (8) and (9).

[0153] In the large-screen field, novel media assets refer to media assets that users have not seen in the last 7 days, clicked in the last 30 days, or played in the last 90 days. Based on the above time windows, a novelty status tag is generated for the media asset to be recommended. If the media asset to be recommended has not been viewed by the target user in the last 7 days, it is marked as "1" in the "not seen in the last 7 days" field; if it has not been clicked in the last 30 days, it is marked as "1" in the "not clicked in the last 30 days" field; similarly, if it has not been played in the last 90 days, it is marked as "1" in the "not played in the last 90 days" field. The above novelty status tag is concatenated with the first media asset feature representation of the media asset to be recommended, and the second media asset feature representation E of the media asset to be recommended is obtained using formula (10). i ′.

[0154] The second media asset characteristic of the media asset to be recommended is represented by E. i The input to the SENet model of the media asset novelty representation module compresses the media asset features of different channels in the media asset to be recommended, that is, compresses each high-dimensional feature into a low-dimensional vector to obtain global information. After compressing the media asset features, these features need to be "incentivized", that is, their weights are recalibrated, and a weight vector with the same dimension as the input is output, representing the importance of different features in evaluating the novelty of the media asset. The weight vector is then used to weight the second media asset feature representation to obtain the third media asset feature representation of the media asset to be recommended. See the following formula for the specific process:

[0155] E″ i =F sg (E′ i (20)

[0156] s i =F ex (E″ i )=f ex，2 (W ex，2 f ex，1 (W ex，1 E″ i)) (twenty one)

[0157] E si =s i ·E″ i (twenty two)

[0158] Among them, F sq This indicates a compression operation, specifically average pooling, where F... ex The excitation operation consists of two parts, namely f ex，1 and f ex，2 They are both feedforward neural networks, but they differ in their activation functions: the former uses ReLU, while the latter uses sigmoid. ex，1 and W ex，2 s represents the trainable parameters. i E represents the weight vector. si This indicates the characteristics of third-party media assets.

[0159] The novelty intent score of the target user calculated using formula (6) is used to weight the feature representation of the third media asset to obtain the novelty representation of the target user for the recommended media asset. The specific process is shown in the following formula:

[0160] E s =p n ·E si (twenty three)

[0161] Among them, E s This indicates the novelty of the recommended media assets to the target users.

[0162] In this implementation, the SENet model introduces an attention mechanism, adjusting the second media asset feature representation through network compression and activation to strengthen key features. This ensures that important features, potentially more attractive to users, are prioritized during the recommendation process. The adjusted third media asset feature representation is then weighted using the target user's novelty intent score, further incorporating the user's immediate novelty needs or interests. Thus, the final media asset novelty representation not only encompasses objective media asset features but also specifically adapts to users' psychological needs. This deep integration of data allows the media recommendation model to more accurately predict which media assets are more likely to provoke user clicks.

[0163] Optionally, determining the predicted click-through rate of the target user for the media asset to be recommended based on the target user's novelty intent representation, the target user's multi-interest representation, and the target user's novelty representation of the media asset to be recommended includes:

[0164] The multi-interest representation of the target user and the novelty representation of the target user for the media asset to be recommended are concatenated and then input into the MLP network of the media asset recommendation model to obtain the fused information representation;

[0165] The novelty representation of the target user to the media asset to be recommended is input into the MLP network of the media asset recommendation model, and the novelty score of the target user to the media asset to be recommended is calculated using the third activation function.

[0166] The novelty score of the target user for the media asset to be recommended is used to calibrate the fused information representation, and the calibrated fused information representation is input into the MLP network of the media asset recommendation model. The predicted click-through rate of the target user for the media asset to be recommended is calculated using the third activation function.

[0167] In one implementation, the target user's multi-interest representation and the target user's novelty representation of the media asset to be recommended are concatenated and input into the MLP network of the media asset recommendation model to obtain the fused information representation E. K E represents the novelty of the recommended media assets as perceived by the target users. s The input is fed into the MLP network of the media asset recommendation model, and the novelty score p of the target user for the recommended media asset is calculated using the third activation function (which can be the sigmoid function). ni Then, utilize the novelty score p of the target users' attitude towards recommended media assets. ni The fused information representation E K Perform calibration and represent the calibrated fusion information as E. K p ni The input is fed into the MLP network of the media asset recommendation model, and the predicted click-through rate of the target user for the recommended media asset is calculated using the third activation function (which can be the sigmoid function).

[0168] In related technologies, some novel media assets have limited overall behavioral data, resulting in significantly lower predicted scores. Since recommendation systems tend to prioritize content with higher predicted scores, novel media assets that should have been exposed may be completely excluded from user exposure due to inaccurate initial predicted scores. This situation not only hinders the accumulation of behavioral data for novel media assets but also negatively impacts the user's novelty experience.

[0169] In this implementation, by calculating and incorporating a novelty score, the potential appeal of media assets to be recommended can be identified. Even with limited global behavioral data, this score helps the model better determine whether the media asset has intrinsic value for users to explore. By using the novelty score to calibrate fusion information, the system can dynamically adjust the importance of media assets under different circumstances, thereby adapting to changes in users' demand for diverse content in the short term. This dynamic calibration increases the probability that users see media content perceived as "new."

[0170] Optionally, the media asset recommendation model is trained in the following manner:

[0171] The user information of the training sample users, the historical playback media asset information of the training sample users, the association information of the media asset recommendation requests of the training sample users, and the media asset information of the training sample media are input into the media asset recommendation model to obtain the predicted click-through rate of the training sample users for the training sample media.

[0172] A binary cross-entropy loss function is constructed using the predicted click-through rate of the training sample users on the training sample media assets and the actual click-through rate of the training sample users on the training sample media assets.

[0173] The media asset recommendation model is trained with the goal of minimizing the binary cross-entropy loss function.

[0174] In one implementation, the predicted click-through rate (CTR) of training sample users on training sample media assets is compared with the actual clicks of training sample users on training sample media assets. The binary cross-entropy loss function is used as the loss function during the training process of the media asset recommendation model, and the model is trained with the goal of minimizing this loss function. The specific representation of the binary cross-entropy loss function is as follows:

[0175]

[0176] in, It represents the actual clicks made by the user of the i-th training sample on the training sample media asset; y (i) is the predicted click-through rate of the i-th training sample user for the training sample media asset; log() represents the logarithmic function with base 2; N represents the number of training samples, and the value of N ranges from 500000 to 1000000.

[0177] In this implementation, the binary cross-entropy loss function effectively measures the degree of inconsistency between the predicted probabilities and the true labels, prompting the model to improve the prediction accuracy for each sample. During weight updates, the loss function pushes the predicted probabilities as close as possible to the actual labels.

[0178] See Figure 3 , Figure 3 This is a structural diagram of a media asset recommendation device provided in another embodiment of this application. (See diagram below.) Figure 3 As shown, the media asset recommendation device 300 includes:

[0179] The first acquisition module 301 is used to acquire, upon receiving a media asset recommendation request from a target user, the target user's user information, the target user's historical playback media asset information, and the associated information of the target user's media asset recommendation request.

[0180] The second acquisition module 302 is used to acquire media asset information of the media assets to be recommended;

[0181] The third acquisition module 303 is used to input the user information of the target user, the historical playback media asset information of the target user, the association information of the media asset recommendation request of the target user, and the media asset information of the media asset to be recommended into the media asset recommendation model, and to obtain the predicted click-through rate of the target user for the media asset to be recommended output by the media asset recommendation model. The media asset recommendation model is obtained by training a neural network model using the user information of multiple training sample users, the historical playback media asset information of the multiple training sample users, the association information of the media asset recommendation request of the multiple training sample users, the media asset information of multiple training sample media, and the actual click situation of each training sample user for each training sample media.

[0182] The first recommendation module 304 is used to determine the target media asset to be recommended to the target user from the media assets to be recommended based on the predicted click-through rate of the target user on the media assets to be recommended, and to recommend the target media asset.

[0183] Optionally, the media asset recommendation model includes a user novelty intent recognition module, a user multi-interest extraction module, and a media asset novelty representation module;

[0184] The third acquisition module 303 includes:

[0185] The first processing unit is used to input the user information of the target user and the association information of the target user's media asset recommendation request into the user novelty intent recognition module to obtain the novelty intent representation of the target user;

[0186] The second processing unit is used to input the target user's historical playback media asset information into the user's multi-interest extraction module to obtain the target user's multi-interest representation;

[0187] The third processing unit is used to input the media asset information of the media asset to be recommended into the media asset novelty representation module to obtain the novelty representation of the media asset to be recommended by the target user;

[0188] The first determining unit is used to determine the predicted click-through rate of the target user for the media asset to be recommended based on the target user's novelty intent representation, the target user's multi-interest representation, and the target user's novelty representation for the media asset to be recommended.

[0189] Optionally, the first processing unit includes:

[0190] The first processing subunit is used to input the user information of the target user into the first embedding layer of the novelty intent recognition module, and use the multilayer perceptron (MLP) network of the novelty intent recognition module to obtain the user feature representation of the target user.

[0191] The second processing subunit is used to input the association information of the target user's media asset recommendation request into the second embedding layer of the novelty intent recognition module, and use the MLP network of the novelty intent recognition module to obtain the target user's recommendation request behavior representation;

[0192] The third processing subunit is used to input the user feature representation and the recommendation request behavior representation into the MLP network of the novelty intent recognition module, and use the first activation function to calculate the novelty intent representation of the target user.

[0193] Optionally, the device further includes:

[0194] The first calculation module is used to calculate the non-novel intent score of the target user based on the novelty intent representation of the target user using the second activation function;

[0195] The historical playback media asset information includes media asset information of multiple media assets played by the target user within a preset historical time period; the second processing unit includes:

[0196] The fourth processing subunit is used to input the media asset information of multiple media assets played by the target user within a preset historical time period into the embedding layer of the user multi-interest extraction module to obtain the media asset information representation of the multiple media assets.

[0197] The first splicing subunit is used to splice the media asset information representations of the multiple media assets in chronological order to obtain the historical playback behavior sequence representation of the target user;

[0198] The fifth processing subunit is used to input the historical playback behavior sequence representation of the target user into the attention layer of the user multi-interest extraction module to obtain the initial multi-interest representation of the target user;

[0199] The first weighting subunit is used to weight the initial multi-interest representation using the non-novelty intent score to obtain the multi-interest representation of the target user.

[0200] Optionally, the device further includes:

[0201] The first calculation module is used to calculate the novelty intent score of the target user based on the novelty intent representation of the target user using a second activation function;

[0202] The third processing unit includes:

[0203] The sixth processing subunit is used to input the media asset information of the media asset to be recommended into the embedding layer of the media asset novelty representation module to obtain the first media asset feature representation of the media asset to be recommended;

[0204] The second splicing subunit is used to splice the novelty status mark of the media asset to be recommended and the first media asset feature representation of the media asset to be recommended to obtain the second media asset feature representation of the media asset to be recommended. The novelty status mark of the media asset to be recommended is obtained based on the target user's operation behavior on the media asset to be recommended in different historical periods.

[0205] The seventh processing subunit is used to input the second media asset feature representation into the compression and incentive network SENet model of the media asset novelty representation module, and to obtain the weight vector of the media asset information of the media asset to be recommended after compression and incentive of the second media asset feature representation;

[0206] The second weighting subunit is used to weight the second media asset feature representation using the weight vector to obtain the third media asset feature representation of the media asset to be recommended.

[0207] The third weighting subunit is used to weight the third media asset feature representation using the novelty intent score to obtain the novelty representation of the media asset to be recommended by the target user.

[0208] Optionally, the first determining unit includes:

[0209] The eighth processing subunit is used to concatenate the target user's multi-interest representation and the target user's novelty representation of the media asset to be recommended, and then input them into the MLP network of the media asset recommendation model to obtain the fused information representation;

[0210] The ninth processing subunit is used to input the novelty representation of the target user to the media asset to be recommended into the MLP network of the media asset recommendation model, and use the third activation function to calculate the novelty score of the target user to the media asset to be recommended.

[0211] The tenth processing subunit is used to calibrate the fused information representation using the novelty score of the target user for the media asset to be recommended, and input the calibrated fused information representation into the MLP network of the media asset recommendation model, and use the third activation function to calculate the predicted click-through rate of the target user for the media asset to be recommended.

[0212] Optionally, the media asset recommendation model is trained in the following manner:

[0213] The user information of the training sample users, the historical playback media asset information of the training sample users, the association information of the media asset recommendation requests of the training sample users, and the media asset information of the training sample media are input into the media asset recommendation model to obtain the predicted click-through rate of the training sample users for the training sample media.

[0214] A binary cross-entropy loss function is constructed using the predicted click-through rate of the training sample users on the training sample media assets and the actual click-through rate of the training sample users on the training sample media assets.

[0215] The media asset recommendation model is trained with the goal of minimizing the binary cross-entropy loss function.

[0216] This application also provides an electronic device. Since the principle by which the electronic device solves the problem is similar to the media asset recommendation method in this application, the implementation of this electronic device can be found in the implementation of the method, and repeated details will not be described again. Figure 4 As shown, the electronic device according to an embodiment of this application includes: a processor 400, configured to read a program from a memory 420 and execute the following processes:

[0217] Upon receiving a media asset recommendation request from a target user, the system obtains the target user's user information, the target user's historical media asset information, and the associated information of the target user's media asset recommendation request.

[0218] Obtain media asset information for media assets to be recommended;

[0219] The target user's user information, the target user's historical playback media asset information, the target user's media asset recommendation request association information, and the media asset information of the media asset to be recommended are input into the media asset recommendation model. The predicted click-through rate of the target user for the media asset to be recommended is obtained from the output of the media asset recommendation model. The media asset recommendation model is obtained by training a neural network model using the user information of multiple training sample users, the historical playback media asset information of the multiple training sample users, the association information of the media asset recommendation requests of the multiple training sample users, the media asset information of multiple training sample media, and the actual click situation of each training sample user for each training sample media.

[0220] Based on the predicted click-through rate of the target user for the media asset to be recommended, a target media asset is determined from the media assets to be recommended and recommended to the target user.

[0221] Optionally, the media asset recommendation model includes a user novelty intent recognition module, a user multi-interest extraction module, and a media asset novelty representation module;

[0222] The processor 400 is used to read the program in the memory 420 and execute the following processes:

[0223] The user information of the target user and the association information of the target user's media asset recommendation request are input into the user novelty intent recognition module to obtain the novelty intent representation of the target user;

[0224] The target user's historical media asset information is input into the user's multi-interest extraction module to obtain the target user's multi-interest representation;

[0225] The media asset information of the media asset to be recommended is input into the media asset novelty representation module to obtain the novelty representation of the media asset to be recommended by the target user;

[0226] Based on the target user's novelty intent representation, the target user's multi-interest representation, and the target user's novelty representation of the media asset to be recommended, the predicted click-through rate of the target user for the media asset to be recommended is determined.

[0227] Optionally, the processor 400 is configured to read the program from the memory 420 and execute the following processes:

[0228] The user information of the target user is input into the first embedding layer of the novelty intent recognition module, and the user feature representation of the target user is obtained by using the multilayer perceptron (MLP) network of the novelty intent recognition module.

[0229] The association information of the target user's media asset recommendation request is input into the second embedding layer of the novelty intent recognition module, and the MLP network of the novelty intent recognition module is used to obtain the target user's recommendation request behavior representation;

[0230] The user feature representation and the recommendation request behavior representation are input into the MLP network of the novelty intent recognition module, and the novelty intent representation of the target user is calculated using the first activation function.

[0231] Optionally, the processor 400 is configured to read the program from the memory 420 and execute the following processes:

[0232] Based on the novelty intent representation of the target user, the non-novelty intent score of the target user is calculated using the second activation function;

[0233] The historical playback media asset information includes media asset information of multiple media assets played by the target user within a preset historical time period; the processor 400 is used to read the program in the memory 420 and execute the following processes:

[0234] The media asset information of multiple media assets played by the target user within a preset historical time period is input into the embedding layer of the user multi-interest extraction module to obtain the media asset information representation of the multiple media assets.

[0235] The media asset information of the multiple media assets is concatenated in chronological order to obtain the historical playback behavior sequence of the target user;

[0236] The historical playback behavior sequence of the target user is input into the attention layer of the user multi-interest extraction module to obtain the initial multi-interest representation of the target user;

[0237] The initial multi-interest representation is weighted using the non-novelty intent score to obtain the multi-interest representation of the target user.

[0238] Optionally, the processor 400 is configured to read the program from the memory 420 and execute the following processes:

[0239] Based on the novelty intent representation of the target user, the novelty intent score of the target user is calculated using the second activation function;

[0240] The processor 400 is used to read the program in the memory 420 and execute the following processes:

[0241] The media asset information of the media asset to be recommended is input into the embedding layer of the media asset novelty representation module to obtain the first media asset feature representation of the media asset to be recommended;

[0242] The novelty status marker of the media asset to be recommended and the first media asset feature representation of the media asset to be recommended are concatenated to obtain the second media asset feature representation of the media asset to be recommended. The novelty status marker of the media asset to be recommended is obtained based on the target user's operation behavior on the media asset to be recommended in different historical periods.

[0243] The second media asset feature representation is input into the compression and incentive network SENet model of the media asset novelty representation module. After compression and incentive, the weight vector of the media asset information of the media asset to be recommended is obtained.

[0244] The second media asset feature representation is weighted using the weight vector to obtain the third media asset feature representation of the media asset to be recommended;

[0245] The novelty intent score is used to weight the feature representation of the third media asset to obtain the novelty representation of the media asset to be recommended by the target user.

[0246] Optionally, the processor 400 is configured to read the program from the memory 420 and execute the following processes:

[0247] The multi-interest representation of the target user and the novelty representation of the target user for the media asset to be recommended are concatenated and then input into the MLP network of the media asset recommendation model to obtain the fused information representation;

[0248] The novelty representation of the target user to the media asset to be recommended is input into the MLP network of the media asset recommendation model, and the novelty score of the target user to the media asset to be recommended is calculated using the third activation function.

[0249] The novelty score of the target user for the media asset to be recommended is used to calibrate the fused information representation, and the calibrated fused information representation is input into the MLP network of the media asset recommendation model. The predicted click-through rate of the target user for the media asset to be recommended is calculated using the third activation function.

[0250] Optionally, the media asset recommendation model is trained in the following manner:

[0251] The user information of the training sample users, the historical playback media asset information of the training sample users, the association information of the media asset recommendation requests of the training sample users, and the media asset information of the training sample media are input into the media asset recommendation model to obtain the predicted click-through rate of the training sample users for the training sample media.

[0252] A binary cross-entropy loss function is constructed using the predicted click-through rate of the training sample users on the training sample media assets and the actual click-through rate of the training sample users on the training sample media assets.

[0253] The media asset recommendation model is trained with the goal of minimizing the binary cross-entropy loss function.

[0254] Among them, Figure 4 In this context, the bus architecture can include any number of interconnected buses and bridges, specifically linking various circuits of one or more processors represented by processor 400 and memory represented by memory 420 together. The bus architecture can also link various other circuits such as peripheral devices, voltage regulators, and power management circuits, which are well known in the art and therefore will not be described further herein. The bus interface provides the interface.

[0255] This application also provides a computer-readable storage medium storing a computer program. When executed by a processor, the computer program implements the various processes of the above-described media resource recommendation method embodiments and achieves the same technical effects. To avoid repetition, it will not be described again here. The computer-readable storage medium may be a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk, etc.

[0256] This application also provides a computer program product, including computer instructions, which, when executed by a processor, implement the above-described... Figure 1 The various processes of the method embodiments shown can achieve the same technical effect, and will not be described again here to avoid repetition.

[0257] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0258] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal (which may be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0259] The embodiments of this application have been described above with reference to the accompanying drawings. However, this application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of this application.

Claims

1. A media asset recommendation method, characterized in that, The method includes: Upon receiving a media asset recommendation request from a target user, the system obtains the target user's user information, the target user's historical media asset information, and the associated information of the target user's media asset recommendation request. Obtain media asset information for media assets to be recommended; The target user's user information, the target user's historical playback media asset information, the target user's media asset recommendation request association information, and the media asset information of the media asset to be recommended are input into the media asset recommendation model. The predicted click-through rate of the target user for the media asset to be recommended is obtained from the output of the media asset recommendation model. The media asset recommendation model is obtained by training a neural network model using the user information of multiple training sample users, the historical playback media asset information of the multiple training sample users, the association information of the media asset recommendation requests of the multiple training sample users, the media asset information of multiple training sample media, and the actual click situation of each training sample user for each training sample media. Based on the predicted click-through rate of the target user for the media asset to be recommended, a target media asset is determined from the media assets to be recommended and recommended to the target user.

2. The media asset recommendation method according to claim 1, characterized in that, The media asset recommendation model includes a user novelty intent recognition module, a user multi-interest extraction module, and a media asset novelty representation module; The step of inputting the target user's user information, the target user's historical media asset information, the association information of the target user's media asset recommendation request, and the media asset information of the media asset to be recommended into the media asset recommendation model, and obtaining the predicted click-through rate of the target user for the media asset to be recommended output by the media asset recommendation model, includes: The user information of the target user and the association information of the target user's media asset recommendation request are input into the user novelty intent recognition module to obtain the novelty intent representation of the target user; The target user's historical media asset information is input into the user's multi-interest extraction module to obtain the target user's multi-interest representation; The media asset information of the media asset to be recommended is input into the media asset novelty representation module to obtain the novelty representation of the media asset to be recommended by the target user; Based on the target user's novelty intent representation, the target user's multi-interest representation, and the target user's novelty representation of the media asset to be recommended, the predicted click-through rate of the target user for the media asset to be recommended is determined.

3. The media asset recommendation method according to claim 2, characterized in that, The step of inputting the user information of the target user and the association information of the target user's media asset recommendation request into the user novelty intent recognition module to obtain the novelty intent representation of the target user includes: The user information of the target user is input into the first embedding layer of the novelty intent recognition module, and the user feature representation of the target user is obtained by using the multilayer perceptron (MLP) network of the novelty intent recognition module. The association information of the target user's media asset recommendation request is input into the second embedding layer of the novelty intent recognition module, and the MLP network of the novelty intent recognition module is used to obtain the target user's recommendation request behavior representation; The user feature representation and the recommendation request behavior representation are input into the MLP network of the novelty intent recognition module, and the novelty intent representation of the target user is calculated using the first activation function.

4. The media asset recommendation method according to claim 2, characterized in that, After inputting the user information of the target user and the association information of the target user's media asset recommendation request into the user novelty intent recognition module to obtain the novelty intent representation of the target user, the method further includes: Based on the novelty intent representation of the target user, the non-novelty intent score of the target user is calculated using the second activation function; The historical playback media asset information includes media asset information of multiple media assets played by the target user within a preset historical time period; the step of inputting the target user's historical playback media asset information into the user multi-interest extraction module to obtain the target user's multi-interest representation includes: The media asset information of multiple media assets played by the target user within a preset historical time period is input into the embedding layer of the user multi-interest extraction module to obtain the media asset information representation of the multiple media assets. The media asset information of the multiple media assets is concatenated in chronological order to obtain the historical playback behavior sequence of the target user; The historical playback behavior sequence of the target user is input into the attention layer of the user multi-interest extraction module to obtain the initial multi-interest representation of the target user; The initial multi-interest representation is weighted using the non-novelty intent score to obtain the multi-interest representation of the target user.

5. The media asset recommendation method according to claim 2, characterized in that, After inputting the user information of the target user and the association information of the target user's media asset recommendation request into the user novelty intent recognition module to obtain the novelty intent representation of the target user, the method further includes: Based on the novelty intent representation of the target user, the novelty intent score of the target user is calculated using the second activation function; The step of inputting the media asset information of the media asset to be recommended into the media asset novelty representation module to obtain the novelty representation of the media asset to be recommended by the target user includes: The media asset information of the media asset to be recommended is input into the embedding layer of the media asset novelty representation module to obtain the first media asset feature representation of the media asset to be recommended; The novelty status marker of the media asset to be recommended and the first media asset feature representation of the media asset to be recommended are concatenated to obtain the second media asset feature representation of the media asset to be recommended. The novelty status marker of the media asset to be recommended is obtained based on the target user's operation behavior on the media asset to be recommended in different historical periods. The second media asset feature representation is input into the compression and incentive network SENet model of the media asset novelty representation module. After compression and incentive, the weight vector of the media asset information of the media asset to be recommended is obtained. The second media asset feature representation is weighted using the weight vector to obtain the third media asset feature representation of the media asset to be recommended; The novelty intent score is used to weight the feature representation of the third media asset to obtain the novelty representation of the media asset to be recommended by the target user.

6. The media asset recommendation method according to claim 2, characterized in that, The step of determining the predicted click-through rate of the target user for the media asset to be recommended based on the target user's novelty intent representation, the target user's multi-interest representation, and the target user's novelty representation for the media asset to be recommended includes: The multi-interest representation of the target user and the novelty representation of the target user for the media asset to be recommended are concatenated and then input into the MLP network of the media asset recommendation model to obtain the fused information representation; The novelty representation of the target user to the media asset to be recommended is input into the MLP network of the media asset recommendation model, and the novelty score of the target user to the media asset to be recommended is calculated using the third activation function. The novelty score of the target user for the media asset to be recommended is used to calibrate the fused information representation, and the calibrated fused information representation is input into the MLP network of the media asset recommendation model. The predicted click-through rate of the target user for the media asset to be recommended is calculated using the third activation function.

7. The media asset recommendation method according to claim 1, characterized in that, The media asset recommendation model was trained in the following manner: The user information of the training sample users, the historical playback media asset information of the training sample users, the association information of the media asset recommendation requests of the training sample users, and the media asset information of the training sample media are input into the media asset recommendation model to obtain the predicted click-through rate of the training sample users for the training sample media. A binary cross-entropy loss function is constructed using the predicted click-through rate of the training sample users on the training sample media assets and the actual click-through rate of the training sample users on the training sample media assets. The media asset recommendation model is trained with the goal of minimizing the binary cross-entropy loss function.

8. A media asset recommendation device, characterized in that, The device includes: The first acquisition module is used to acquire, upon receiving a media asset recommendation request from a target user, the target user's user information, the target user's historical playback media asset information, and the associated information of the target user's media asset recommendation request. The second acquisition module is used to acquire media asset information of the media assets to be recommended; The third acquisition module is used to input the user information of the target user, the historical playback media asset information of the target user, the association information of the media asset recommendation request of the target user, and the media asset information of the media asset to be recommended into the media asset recommendation model, and to obtain the predicted click-through rate of the target user for the media asset to be recommended output by the media asset recommendation model. The media asset recommendation model is obtained by training a neural network model using the user information of multiple training sample users, the historical playback media asset information of the multiple training sample users, the association information of the media asset recommendation request of the multiple training sample users, the media asset information of multiple training sample media, and the actual click situation of each training sample user for each training sample media. The first recommendation module is used to determine the target media asset to be recommended to the target user from the media assets to be recommended based on the predicted click-through rate of the target user on the media assets to be recommended, and to recommend the target media asset.

9. An electronic device, characterized in that, The electronic device includes a processor, the processor being configured to: Upon receiving a media asset recommendation request from a target user, the system obtains the target user's user information, the target user's historical media asset information, and the associated information of the target user's media asset recommendation request. Obtain media asset information for media assets to be recommended; The target user's user information, the target user's historical playback media asset information, the target user's media asset recommendation request association information, and the media asset information of the media asset to be recommended are input into the media asset recommendation model. The predicted click-through rate of the target user for the media asset to be recommended is obtained from the output of the media asset recommendation model. The media asset recommendation model is obtained by training a neural network model using the user information of multiple training sample users, the historical playback media asset information of the multiple training sample users, the association information of the media asset recommendation requests of the multiple training sample users, the media asset information of multiple training sample media, and the actual click situation of each training sample user for each training sample media. Based on the predicted click-through rate of the target user for the media asset to be recommended, a target media asset is determined from the media assets to be recommended and recommended to the target user.

10. An electronic device, characterized in that, It includes a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the steps of the media asset recommendation method as described in any one of claims 1 to 7.

11. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the media asset recommendation method as described in any one of claims 1 to 7.

12. A computer program product, characterized in that, Includes computer instructions that, when executed by a processor, implement the steps of the method as described in any one of claims 1 to 7.

Images