Training Method, Device and Computer Equipment for User Preference Prediction Model

Abstract

The present application relates to a method, device, and computer equipment for training a user preference prediction model. The present application relates to the field of big data technology. The method includes: obtaining historical browsing sample information of users; each piece of historical browsing sample information includes historical click browsing sample information of each user and historical preference browsing sample information of the user; through the convolutional layer of the initial user preference prediction model, performing dilated convolution processing on each piece of historical click browsing sample information, and through the extraction layer of the initial user preference prediction model, extracting click feature information of the historical click browsing sample information after each dilated convolution processing; through the prediction layer of the initial user preference prediction model, predicting the predicted preference information of the user according to the extracted click feature information, and training the initial user preference prediction model to obtain a user preference prediction model. Using this method can improve the training efficiency of the user preference prediction model.

Description

Technical Field

[0001] This application relates to the field of big data technology, and particularly to a method, apparatus, and computer device for training a user preference prediction model. Background Art

[0002] As the basis of a large number of applications, the user preference prediction system can help applications capture users' interest preferences well and recommend content that users are interested in to users. Therefore, it has become increasingly important to effectively improve the performance of the user preference prediction system.

[0003] In the existing user preference prediction model, multiple categorical variables are mapped into an abstract real-valued space through a deep neural network to extract high-dimensional user preference feature information for each category. And the prediction network is used to predict each user preference feature information to obtain the predicted user preference information. However, the disadvantage of the existing technology is that the high-dimensional data generated by the deep neural network increases the training time cost of the user preference prediction model, resulting in a low training efficiency of the user preference prediction model. Summary of the Invention

[0004] Based on this, in view of the above technical problems, it is necessary to provide a method, apparatus, computer device, computer-readable storage medium, and computer program product for training a user preference prediction model.

[0005] In a first aspect, this application provides a method for training a user preference prediction model. The method includes:

[0006] Obtain historical browsing sample information of users; each piece of the historical browsing sample information includes historical click-browsing sample information of each user and historical preference-browsing sample information of the user;

[0007] Perform dilated convolution processing on each piece of the historical click-browsing sample information through the convolutional layer of the initial user preference prediction model, and extract click feature information of the historical click-browsing sample information after each dilated convolution processing through the extraction layer of the initial user preference prediction model;

[0008] Through the prediction layer of the initial user preference prediction model, predict the predicted preference information of the user according to the extracted click feature information, and train the initial user preference prediction model according to the preference-browsing sample information of the user and the predicted preference information of the user to obtain a user preference prediction model.

[0009] Optionally, the training the initial user preference prediction model according to the preference-browsing sample information of the user and the predicted preference information of the user to obtain a user preference prediction model includes:

[0010] Calculate the similarity between the predicted preference information and the preference browsing sample information through a loss function;

[0011] In the case where the similarity between the predicted preference information and the preference browsing sample information is less than the similarity threshold, update the weight parameters of the prediction layer and the extraction parameters of the extraction layer, and return to execute the step of performing dilated convolution processing on each of the historical click-browse sample information through the convolutional layer of the initial user preference prediction model until the similarity between the predicted preference information and the preference browsing sample information is greater than the similarity threshold;

[0012] Use the initial user preference prediction model corresponding to the predicted preference information greater than the similarity threshold as the user preference prediction model.

[0013] Optionally, before performing dilated convolution processing on each of the historical click-browse sample information through the convolutional layer of the initial user preference prediction model and extracting the click feature information of the historical click-browse sample information after each dilated convolution processing through the extraction layer of the initial user preference prediction model, it further includes:

[0014] Extract the click feature information of each of the historical click-browse sample information through the extraction layer of the initial user preference prediction model.

[0015] Optionally, performing dilated convolution processing on each of the historical click-browse sample information through the convolutional layer of the initial user preference prediction model and extracting the click feature information of the historical click-browse sample information after each dilated convolution processing through the extraction layer of the initial user preference prediction model includes:

[0016] Perform dilated convolution processing on each of the historical click-browse sample information through the convolutional layer of the initial user preference prediction model, and extract the click feature information of the dilated-convolved historical click-browse sample information through the extraction layer;

[0017] In the case where the number of dilated convolutions has not reached the preset number of dilated convolutions, input the dilated-convolved historical click-browse sample information into the convolutional layer, and return to execute the step of performing dilated convolution processing on each of the historical click-browse sample information through the convolutional layer of the initial user preference prediction model and extracting the click feature information of the dilated-convolved historical click-browse sample information through the extraction layer until the number of dilated convolutions reaches the preset number of dilated convolutions, and output the click feature information extracted each time.

[0018] Optionally, the click and browse sample information includes the sample click and browse rate. Predicting the predicted preference information of the user according to each extracted click feature information through the prediction layer of the initial user preference prediction model includes:

[0019] Calculating the sample weight value corresponding to each click and browse sample information according to the sample click and browse rate of each click and browse sample information, and using the weight value corresponding to each click and browse sample information as the weight value of each click feature information extracted through the click and browse sample information in the prediction layer of the initial user preference prediction model;

[0020] Performing weighted processing on each click feature information according to the weight value of each click feature information to obtain each target click feature information;

[0021] According to each target click feature information, through the prediction layer of the initial user preference prediction model, screening the target historical browse sample information corresponding to each target click feature information in each historical browse sample information, and using the target historical browse sample information as the predicted preference information of the user.

[0022] In a second aspect, the present application provides a method for a user preference prediction model. The method includes:

[0023] Obtaining the historical click and browse information of the target user and the candidate information to be browsed by the target user;

[0024] Performing dilated convolution processing on the historical click and browse information through the convolution layer of the user preference prediction model, and extracting the click feature information of the click and browse information after each dilated convolution processing through the extraction layer of the user preference prediction model;

[0025] Predicting the target preference browse information of the target user in each candidate information according to each click feature information through the prediction layer of the user preference prediction model;

[0026] Wherein, the user preference prediction model is trained by the training method of the user preference prediction model according to any one of the first aspects.

[0027] In a third aspect, the present application further provides a training device for a user preference prediction model. The device includes:

[0028] An acquisition module, configured to acquire the user's multiple historical browse sample information; each historical browse sample information includes the user's historical click and browse sample information and the user's historical preference browse sample information;

[0029] The first feature extraction module is used to perform dilated convolution processing on each of the historical click and browse sample information through the convolutional layer of the initial user preference prediction model, and extract the click feature information of the click and browse sample information after each dilated convolution processing through the extraction layer of the initial user preference prediction model;

[0030] The training module is used to predict the predicted preference information of the user according to the extracted click feature information through the prediction layer of the initial user preference prediction model, and train the initial user preference prediction model according to the preference browsing sample information of the user and the predicted preference information of the user to obtain a user preference prediction model.

[0031] Optionally, the training module is specifically used for:

[0032] Calculate the similarity between the predicted preference information and the preference browsing sample information through a loss function;

[0033] In the case where the similarity between the predicted preference information and the preference browsing sample information is less than the similarity threshold, update the weight parameters of the prediction layer and the extraction parameters of the extraction layer, and return to execute the step of performing dilated convolution processing on each of the historical click and browse sample information through the convolutional layer of the initial user preference prediction model until the similarity between the predicted preference information and the preference browsing sample information is greater than the similarity threshold;

[0034] Use the initial user preference prediction model corresponding to the predicted preference information greater than the similarity threshold as the user preference prediction model.

[0035] Optionally, the device further includes:

[0036] The second feature extraction module is used to extract the click feature information of each of the historical click and browse sample information through the extraction layer of the initial user preference prediction model.

[0037] Optionally, the first feature extraction module is specifically used for:

[0038] Perform dilated convolution processing on each of the historical click and browse sample information through the convolutional layer of the initial user preference prediction model, and extract the click feature information of the historical click and browse sample information that has undergone dilated convolution processing through the extraction layer;

[0039] In the case where the number of dilated convolutions has not reached the preset number of dilated convolutions, input the historical click-browse sample information that has been dilated-convolved into the convolutional layer, and return to execute the convolutional layer of the initial user preference prediction model to perform dilated convolution processing on each piece of the historical click-browse sample information, and through the extraction layer, extract the click feature information of the historical click-browse sample information that has been dilated-convolved, until the number of dilated convolutions reaches the preset number of dilated convolutions, and output the click feature information extracted each time.

[0040] Optionally, the training module is specifically configured to:

[0041] Calculate the sample weight value corresponding to each click-browse sample information according to the sample click-browse rate of each click-browse sample information, and use the weight value corresponding to each click-browse sample information as the weight value of each click feature information extracted through the click-browse sample information in the prediction layer of the initial user preference prediction model;

[0042] Perform weighted processing on each click feature information according to the weight value of each click feature information to obtain each target click feature information;

[0043] According to each target click feature information, through the prediction layer of the initial user preference prediction model, screen the target historical browse sample information corresponding to each target click feature information in each historical browse sample information, and use the target historical browse sample information as the predicted preference information of the user.

[0044] In a fourth aspect, the present application also provides a user preference prediction model device. The device includes:

[0045] A target information acquisition module, configured to acquire the historical click-browse information of a target user and the candidate information to be browsed by the target user;

[0046] A feature extraction module, configured to perform dilated convolution processing on the historical click-browse information through the convolutional layer of the user preference prediction model, and extract the click feature information of the click-browse information after each dilated convolution processing through the extraction layer of the user preference prediction model;

[0047] A prediction module, configured to predict the target preference browse information of the target user in each piece of the candidate information according to each click feature information through the prediction layer of the user preference prediction model.

[0048] Wherein, the user preference prediction model is trained by the training method of the user preference prediction model described in any one of the first aspects.

[0049] Fifth aspect, the present application provides a computer device. The computer device includes a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, the steps of the method according to any one of the first aspect or the second aspect are implemented.

[0050] Sixth aspect, the present application provides a computer-readable storage medium. A computer program is stored thereon, and when the computer program is executed by a processor, the steps of the method according to any one of the first aspect or the second aspect are implemented.

[0051] Seventh aspect, the present application provides a computer program product. The computer program product includes a computer program, and when the computer program is executed by a processor, the steps of the method according to any one of the first aspect or the second aspect are implemented.

[0052] The above-mentioned training method, device and computer device of the user preference prediction model obtain user's multiple historical browsing sample information; each of the historical browsing sample information includes each user's historical click-browsing sample information and the user's historical preference-browsing sample information; through the convolutional layer of the initial user preference prediction model, perform dilated convolution processing on each of the historical click-browsing sample information, and through the extraction layer of the initial user preference prediction model, extract the click feature information of the click-browsing sample information after each dilated convolution processing; through the prediction layer of the initial user preference prediction model, according to each of the click feature information, predict the predicted preference information of the user, and according to the user's preference-browsing sample information and the user's predicted preference information, train the initial user preference prediction model to obtain the user preference prediction model. By performing dilated convolution processing on each historical click-browsing sample information, the dimension of each click-browsing sample information is reduced, and by performing feature extraction operations on the click-browsing sample information after each dilated convolution processing, each click feature information is obtained, so as to reduce the dimension of the data (i.e., the click-browsing sample information) while ensuring high-fidelity information extraction, and improve the training efficiency of the user preference prediction model. Description of the Drawings

[0053] Figure 1 It is a schematic flowchart of the training method of the user preference prediction model in an embodiment;

[0054] Figure 2 It is a schematic flowchart of the extraction steps of the click feature information in an embodiment;

[0055] Figure 3 It is a schematic flowchart of the user preference prediction method in an embodiment;

[0056] Figure 4A flowchart showing an example of training a user preference prediction model in an embodiment;

[0057] Figure 5 A structural block diagram of a training device for a user preference prediction model in an embodiment;

[0058] Figure 6 A structural block diagram of a user preference prediction device in an embodiment;

[0059] Figure 7 An internal structural diagram of a computer device in an embodiment. Detailed implementation manners

[0060] In order to make the objectives, technical solutions and advantages of the present application more clear and understandable, the present application will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0061] The training method for the user preference prediction model provided in the embodiments of the present application, and the training method for the test result classification model provided in the embodiments of the present application can be applied to a terminal, or to a server, or to a system including a terminal and a server, and is implemented through the interaction between the terminal and the server. Among them, the terminal may include, but is not limited to, various personal computers, laptop computers, tablet computers, etc. The terminal performs dilated convolution processing on each historical click-browse sample information to reduce the dimension of each click-browse sample information, and performs feature extraction operations on the click-browse sample information obtained by each dilated convolution processing to obtain each click feature information, so as to reduce the dimension of the data (i.e., the click-browse sample information) while ensuring high-fidelity information extraction, and improve the training efficiency of the user preference prediction model.

[0062] In one embodiment, as Figure 1 shown, a training method for a user preference prediction model is provided. Taking the application of this method to a terminal as an example, the method includes the following steps:

[0063] Step S101, obtain the user's multiple historical browsing sample information.

[0064] Among them, each historical browsing sample information includes the historical click-browse sample information of each user and the historical preference browsing sample information of the user.

[0065] In this embodiment, the terminal obtains the historical browsing records of the user in the client based on the user's authorized operation, and sorts the browsing addresses in each historical browsing record in descending order of the browsing times. The terminal selects the browsing address with the most browsing times as the historical preference browsing sample address, and uses the content corresponding to this address as the historical preference browsing sample information. The terminal uses the content corresponding to the other historical browsing records as the respective click browsing sample information. Among them, the client can be, but is not limited to, an independently operated client type such as a web client, an application, or a mini-program.

[0066] Step S102: Perform dilated convolution processing on each piece of historical click browsing sample information through the convolutional layer of the initial user preference prediction model, and extract the click feature information of the historical click browsing sample information after each dilated convolution processing through the extraction layer of the initial user preference prediction model.

[0067] In this embodiment, the terminal performs dilated convolution processing on each piece of historical click browsing sample information through the convolutional layer of the initial user preference prediction model to reduce the dimension of the historical click browsing sample information. Then, the terminal extracts the click feature information of the historical click browsing sample information after the dilated convolution processing through the extraction layer of the initial user preference prediction model after each dilated convolution processing, obtaining multiple pieces of click feature information. Among them, the convolutional layer is a stacked gated causal convolutional neural network with different dilation rates, and the extraction layer is a feature information extraction network based on the self-attention mechanism.

[0068] Step S103: Through the prediction layer of the initial user preference prediction model, predict the predicted preference information of the user according to the extracted click feature information, and train the initial user preference prediction model according to the user's preference browsing sample information and the user's predicted preference information to obtain the user preference prediction model.

[0069] In this embodiment, the terminal performs weighted summation processing on all the extracted click feature information through the prediction layer of the initial user preference prediction model to obtain the target click feature information. The terminal predicts the predicted preference information of the user according to the obtained target click feature information. The terminal trains the initial user preference prediction model based on the user's preference browsing sample information and the user's predicted preference information to obtain the user preference prediction model.

[0070] Based on the above solution, by performing dilated convolution processing on each piece of historical click browsing sample information, the dimension of each click browsing sample information is reduced, and by performing feature extraction operations on the click browsing sample information after each dilated convolution processing, each piece of click feature information is obtained, thereby reducing the dimension of the data (i.e., the click browsing sample information) while ensuring high-fidelity information extraction, and improving the training efficiency of the user preference prediction model.

[0071] Optionally, according to the user's preference for browsing sample information and the user's predicted preference information, the initial user preference prediction model is trained to obtain a user preference prediction model, including: calculating the similarity between the predicted preference information and the preference browsing sample information through a loss function; in the case where the similarity between the predicted preference information and the preference browsing sample information is less than the similarity threshold, updating the weight parameters of the prediction layer and the extraction parameters of the extraction layer, and returning to execute the step of performing dilated convolution processing on each historical click-browse sample information through the convolutional layer of the initial user preference prediction model until the similarity between the predicted preference information and the preference browsing sample information is greater than the similarity threshold; using the initial user preference prediction model corresponding to the predicted preference information greater than the similarity threshold as the user preference prediction model.

[0072] In this embodiment, the terminal presets a similarity threshold and calculates the similarity between the predicted preference information and the preference browsing sample information through a loss function. The terminal determines the magnitude relationship between the similarity between the predicted preference information and the preference browsing sample information and the similarity threshold. In the case where the similarity between the predicted preference information and the preference browsing sample information is greater than the similarity threshold, the terminal uses the initial user preference prediction model corresponding to the similarity not less than the similarity threshold as the user preference prediction model. In the case where the similarity between the predicted preference information and the preference browsing sample information is less than the similarity threshold, the terminal updates the weight parameters of the prediction layer and the extraction parameters of the extraction layer, and returns to execute step S102 until the similarity between the predicted preference information and the preference browsing sample information is greater than the similarity threshold. The terminal uses the initial user preference prediction model corresponding to the predicted preference information greater than the similarity threshold as the user preference prediction model.

[0073] Based on the above solution, by calculating the similarity between the predicted preference information and the preference browsing sample information, it is possible to determine whether the user preference prediction model is successfully trained, thereby optimizing the accuracy of training the user preference prediction model.

[0074] Optionally, before performing dilated convolution processing on each historical click-browse sample information through the convolutional layer of the initial user preference prediction model and extracting the click feature information of the click-browse sample information after each dilated convolution processing through the extraction layer of the initial user preference prediction model, it further includes:

[0075] Extracting the click feature information of each historical click-browse sample information through the extraction layer of the initial user preference prediction model.

[0076] In this embodiment, before performing dilated convolution processing, the terminal first extracts the click feature information of each historical click-browse sample information through the extraction layer of the initial user preference prediction model to retain the feature degree of each historical click-browse sample information.

[0077] Specifically, the terminal is given an input time series x of a user's historical behavior (i.e., historical click and browse sample information). 1: , where x 1: represents the historical click and browse sample information at each moment from the start moment of the historical behavior to the current moment, sorted in a time series. The terminal first uses a self-attention layer to model the internal step-by-step dependencies. x is the historical click and browse sample information. In the time series, the click feature information of x is obtained by fusing the feature information with high similarity at different moments of x. Among them, the process of fusing vectors can be expressed by the following formula:

[0078]

[0079] In the above formula, Q, K, V are tensors of x at different moments, and K T is the reciprocal of K, d k is the similarity between K and x. By performing a correlation coefficient weighting operation on the feature sequence, a row of fused vectors atten(i) (i.e., the click feature information of each historical click and browse sample information) at the i-th step is obtained. The residual connection is used to update the sequence representation as x i .

[0080] Based on the above scheme, by pre-extracting the click feature information of each historical click and browse sample information, the feature degree of each historical click and browse sample information is retained to the greatest extent, ensuring high fidelity of information extraction and avoiding loss of some feature information due to dimensionality reduction, thereby improving the accuracy of training the user preference prediction model.

[0081] Optionally, as Figure 2 shown, through the convolutional layer of the initial user preference prediction model, dilated convolution processing is performed on each historical click and browse sample information, and through the extraction layer of the initial user preference prediction model, the click feature information of the click and browse sample information after each dilated convolution processing is extracted, including:

[0082] Step S201, through the convolutional layer of the initial user preference prediction model, perform dilated convolution processing on each historical click and browse sample information, and through the extraction layer, extract the click feature information of the historical click and browse sample information that has undergone dilated convolution processing.

[0083] In this embodiment, the terminal performs dilated convolution processing on each historical click and browse sample information respectively through the convolutional layer of the initial user preference prediction model, obtaining multiple historical click and browse sample information that has undergone dilated processing. The terminal extracts the click feature information of the historical click and browse sample information that has undergone dilated convolution processing through the extraction layer.

[0084] Specifically, the dilated convolution processing process is as follows:

[0085] The terminal expands causal convolution and gated linear units in the convolutional layer through dilation. The dilated convolution receives a wider receptive field by skipping the "holes" at intervals (i.e., performs dilated convolution processing on each historical click-browse sample information separately). The convolutional kernel size is 2w + 1. By hierarchically stacking convolutional layers with a larger dilation rate, the historical click-browse sample information expands exponentially, while the number of parameters only grows linearly. In this way, information across multiple-scale time spans can be gradually obtained with fewer layers, which is beneficial to saving computational consumption and avoiding information loss caused by downsampling operations.

[0086] Step S202, when the number of dilated convolutions has not reached the preset number of dilated convolutions, input the historical click-browse sample information that has been processed by dilated convolution into the convolutional layer, and return to execute the convolutional layer of the initial user preference prediction model to perform dilated convolution processing on each historical click-browse sample information, and through the extraction layer, extract the click feature information of the historical click-browse sample information that has been processed by dilated convolution until the number of dilated convolutions reaches the preset number of dilated convolutions, and output the click feature information extracted each time.

[0087] In this embodiment, for each historical click-browse sample information, the terminal presets the number of dilated convolutions and determines whether the number of dilated convolutions has reached the preset number of dilated convolutions. When the number of dilated convolutions has not reached the preset number of dilated convolutions, the terminal inputs the historical click-browse sample information that has been processed by dilated convolution into the convolutional layer, and returns to execute step S201 to continue performing dilated convolution processing on the historical click-browse sample information that has been processed by dilated convolution. When the number of dilated convolutions reaches the preset number of dilated convolutions, the terminal outputs the click feature information extracted each time.

[0088] Based on the above solution, through multiple dilated convolution processes, the historical click-browse sample information is dimensionally reduced multiple times, thereby further improving the training efficiency of the user preference prediction model.

[0089] Optionally, clicking to view sample information includes the sample click-through rate. Through the prediction layer of the initial user preference prediction model, based on the extracted click feature information, the predicted preference information of the user is predicted, including: calculating the sample weight value corresponding to each click-through sample information according to the sample click-through rate of each click-through sample information, and using the weight value corresponding to each click-through sample information as the weight value of each click feature information extracted through the click-through sample information in the prediction layer of the initial user preference prediction model. According to the weight values of the click feature information, the click feature information is weighted to obtain each target click feature information. According to each target click feature information, through the prediction layer of the initial user preference prediction model, in each historical browsing sample information, the target historical browsing sample information corresponding to each target click feature information is screened, and the target historical browsing sample information is used as the predicted preference information of the user.

[0090] In this embodiment, the terminal calculates the sample click-through rate of each click-through sample information according to the sample click count of each click-through sample information and the total click count of all click-through information. The terminal normalizes the sample click-through rate of each click-through sample information, and uses the normalized sample click-through rate as the sample weight value corresponding to the click-through sample information. The terminal uses the weight value corresponding to each click-through sample information as the weight value of each click feature information extracted through the click-through sample information in the prediction layer of the initial user preference prediction model.

[0091] The terminal weights each click feature information according to the weight values of the click feature information to obtain each target click feature information. The terminal, according to each target click feature information, through the prediction layer of the initial user preference prediction model, screens the target historical browsing sample information corresponding to each target click feature information in each historical browsing sample information, and uses the target historical browsing sample information as the predicted preference information of the user.

[0092] Specifically, in the above steps, finally, the target attention is used to adaptively calculate the weights of the stacked representation layers with different feature scales, and the dense representation of each user preference feature is:

[0093]

[0094] After that, all the dense features form a tensor Z. l-L is all the click feature information after numbering the click feature information, and β l is the weight of the click feature information, is the click feature information, and the probability value of the user clicking on a specific item (i.e., the target click feature information) is obtained through the fully connected layer neural network The model parameters are optimized through the cross-entropy loss.

[0095] Based on the above solution, by calculating the weights of click feature information, the predicted preference information of the user is predicted, improving the accuracy of the predicted preference information.

[0096] In one embodiment, as Figure 3 shown, a user preference prediction method is provided. Taking the application of this method to a terminal as an example for illustration, it includes the following steps:

[0097] Step S301, obtain the historical click and browse information of the target user, and the candidate information that the target user is about to browse.

[0098] In this embodiment, based on the user's authorization, the terminal obtains the historical click and browse information of the target user, and the candidate information that the target user is about to browse. For the specific obtaining steps, refer to step S101 for details.

[0099] Step S302, through the convolutional layer of the user preference prediction model, perform dilated convolution processing on the historical click and browse information, and through the extraction layer of the user preference prediction model, extract the click feature information of the click and browse information after each dilated convolution processing.

[0100] In this embodiment, the terminal performs dilated convolution processing on the historical click and browse information through the convolutional layer of the user preference prediction model, and after each dilated convolution processing, through the extraction layer of the user preference prediction model, extracts the click feature information of the click and browse information that has undergone dilated convolution processing until the last dilated convolution processing is completed. The terminal obtains multiple click feature information for each click and browse information. For the specific processing process, refer to step S102 for details.

[0101] Step S303, according to each click feature information, through the prediction layer of the user preference prediction model, predict the target preference browsing information of the target user among the candidate information.

[0102] Among them, the user preference prediction model is trained by the training method of the user preference prediction model according to any one of claims 1 to 5.

[0103] In this embodiment, the terminal calculates the weight value corresponding to each click and browse information according to the click-through rate of each click and browse information, and uses the weight value corresponding to each click and browse information as the weight value of each click feature information extracted through the click and browse information in the prediction layer of the user preference prediction model. According to the weight values of each click feature information, weighted processing is performed on each click feature information to obtain each target click feature information. According to each target click feature information, through the prediction layer of the initial user preference prediction model, among the candidate information, the candidate information corresponding to each target click feature information is screened, and the target historical browsing information is used as the target preference browsing information of the target user.

[0104] Based on the above solution, the terminal predicts the target preference browsing information of the target user through the prediction layer of the trained user preference prediction model, improving the prediction accuracy of the target preference browsing information.

[0105] It should be understood that although each step in the flowchart involved in the above-described embodiments is sequentially shown according to the indication of the arrow, these steps are not necessarily executed in the order indicated by the arrow. Unless there is a clear indication in this article, the execution of these steps has no strict order limit, and these steps can be executed in other orders. Moreover, at least a part of the steps in the flowchart involved in the above-described embodiments may include multiple steps or multiple stages. These steps or stages are not necessarily executed at the same moment, but can be executed at different moments. The execution order of these steps or stages is not necessarily sequential, but can be executed alternately or in turn with at least a part of other steps or steps or stages in other steps.

[0106] This application also provides a training example of a user preference prediction model. As shown in FIG. 4, the specific processing process includes the following steps:

[0107] Step S401, obtain the user's multiple historical browsing sample information.

[0108] Step S402, through the extraction layer of the initial user preference prediction model, extract the click feature information of each historical click browsing sample information.

[0109] Step S403, through the convolutional layer of the initial user preference prediction model, perform dilated convolution processing on each historical click browsing sample information, and through the extraction layer, extract the click feature information of the historical click browsing sample information that has been dilated convolution processed.

[0110] Step S404, when the number of dilated convolutions has not reached the preset number of dilated convolutions, input the historical click browsing sample information that has been dilated convolution processed into the convolutional layer, and return to execute the step of performing dilated convolution processing on each historical click browsing sample information through the convolutional layer of the initial user preference prediction model, and through the extraction layer, extract the click feature information of the historical click browsing sample information that has been dilated convolution processed, until the number of dilated convolutions reaches the preset number of dilated convolutions, and output the click feature information extracted each time.

[0111] Step S405, according to the sample click browsing rate of each click browsing sample information, calculate the sample weight value corresponding to each click browsing sample information, and use the weight value corresponding to each click browsing sample information as the weight value of each click feature information extracted through the click browsing sample information in the prediction layer of the initial user preference prediction model.

[0112] Step S406: According to the weight values of the click feature information, perform weighted processing on the click feature information to obtain respective target click feature information.

[0113] Step S407: According to the respective target click feature information, through the prediction layer of the initial user preference prediction model, in each historical browsing sample information, screen the target historical browsing sample information corresponding to each target click feature information, and use the target historical browsing sample information as the predicted preference information of the user.

[0114] Step S408: Calculate the similarity between the predicted preference information and the preference browsing sample information through a loss function.

[0115] Step S409: In the case where the similarity between the predicted preference information and the preference browsing sample information is less than the similarity threshold, update the weight parameters of the prediction layer and the extraction parameters of the extraction layer, and return to execute the step of performing dilated convolution processing on each historical click browsing sample information through the convolution layer of the initial user preference prediction model until the similarity between the predicted preference information and the preference browsing sample information is greater than the similarity threshold.

[0116] Step S410: Use the initial user preference prediction model corresponding to the predicted preference information greater than the similarity threshold as the user preference prediction model.

[0117] Based on the same inventive concept, an embodiment of the present application further provides a training device for a user preference prediction model for implementing the training method of the user preference prediction model involved above. The implementation solutions provided by this device for solving problems are similar to the implementation solutions described in the above method. Therefore, the specific limitations in one or more embodiments of the training device for the user preference prediction model provided below can refer to the limitations on the training method of the user preference prediction model in the above text, and will not be repeated here.

[0118] In one embodiment, as Figure 5 shown, a training device for a user preference prediction model is provided, including: an acquisition module 510, a first feature extraction module 520, and a training module 530, where:

[0119] The acquisition module 510 is configured to acquire historical browsing sample information of a user in multiple; each of the historical browsing sample information includes historical click browsing sample information of each user and historical preference browsing sample information of the user;

[0120] The first feature extraction module 520 is configured to perform dilated convolution processing on each of the historical click browsing sample information through the convolution layer of the initial user preference prediction model, and extract click feature information of the click browsing sample information after each dilated convolution processing through the extraction layer of the initial user preference prediction model;

[0121] A training module 530, configured to, through a prediction layer of the initial user preference prediction model, predict predicted preference information of the user according to each extracted click feature information, and train the initial user preference prediction model according to the preference browsing sample information of the user and the predicted preference information of the user, so as to obtain a user preference prediction model.

[0122] Optionally, the training module 530 is specifically configured to:

[0123] Calculate a similarity between the predicted preference information and the preference browsing sample information through a loss function;

[0124] In a case where the similarity between the predicted preference information and the preference browsing sample information is less than a similarity threshold, update weight parameters of the prediction layer and extraction parameters of the extraction layer, and return to execute a step of performing dilated convolution processing on each of the historical click browsing sample information through a convolution layer of the initial user preference prediction model until the similarity between the predicted preference information and the preference browsing sample information is greater than the similarity threshold;

[0125] Use the initial user preference prediction model corresponding to the predicted preference information greater than the similarity threshold as the user preference prediction model.

[0126] Optionally, the apparatus further includes:

[0127] A second feature extraction module, configured to extract click feature information of each of the historical click browsing sample information through an extraction layer of the initial user preference prediction model.

[0128] Optionally, the first feature extraction module 520 is specifically configured to:

[0129] Perform dilated convolution processing on each of the historical click browsing sample information through a convolution layer of the initial user preference prediction model, and extract click feature information of the dilated convolution processed historical click browsing sample information through the extraction layer;

[0130] In a case where the number of dilated convolutions does not reach a preset number of dilated convolutions, input the dilated convolution processed historical click browsing sample information into the convolution layer, and return to execute a step of performing dilated convolution processing on each of the historical click browsing sample information through the convolution layer of the initial user preference prediction model, and extracting click feature information of the dilated convolution processed historical click browsing sample information through the extraction layer until the number of dilated convolutions reaches the preset number of dilated convolutions, and output click feature information extracted each time.

[0131] Optionally, the training module 530 is specifically configured to:

[0132] According to the sample click-through rate of each sample of click-through view information, calculate the sample weight value corresponding to each sample of click-through view information, and use the weight value corresponding to each sample of click-through view information as the weight value of each click feature information extracted through the click-through view information in the prediction layer of the initial user preference prediction model;

[0133] Perform weighted processing on each of the click feature information according to the weight value of each of the click feature information to obtain each target click feature information;

[0134] According to each target click feature information, through the prediction layer of the initial user preference prediction model, screen the target historical view sample information corresponding to each target click feature information from each historical view sample information, and use the target historical view sample information as the predicted preference information of the user.

[0135] Based on the same inventive concept, an embodiment of the present application further provides a user preference prediction device for implementing the user preference prediction method involved above. The implementation solution provided by this device to solve the problem is similar to the implementation solution described in the above method. Therefore, the specific limitations in one or more embodiments of the user preference prediction device provided below can refer to the limitations on the user preference prediction method in the above text, and will not be repeated here.

[0136] In one embodiment, as Figure 6 shown, a user preference prediction device is provided, including: a target information acquisition module 610, a feature extraction module 620, and a prediction module 630, where:

[0137] The target information acquisition module 610 is configured to acquire the historical click-through view information of the target user and the candidate information to be browsed by the target user;

[0138] The feature extraction module 620 is configured to perform dilated convolution processing on the historical click-through view information through the convolution layer of the user preference prediction model, and extract the click feature information of the click-through view information after each dilated convolution processing through the extraction layer of the user preference prediction model;

[0139] The prediction module 630 is configured to predict the target preference view information of the target user from each of the candidate information through the prediction layer of the user preference prediction model according to each of the click feature information.

[0140] Wherein, the user preference prediction model is trained by the training method of the user preference prediction model described in any item of the first aspect.

[0141] Each module in the above-mentioned training device and user preference prediction device of the user preference prediction model can be implemented in whole or in part by software, hardware, or a combination thereof. Each of the above modules can be embedded in the processor of the computer device in hardware form or be independent of it, or be stored in the memory of the computer device in software form, so as to facilitate the processor to call and execute the operations corresponding to each of the above modules.

[0142] In one embodiment, a computer device is provided. The computer device can be a terminal, and its internal structural diagram can be as Figure 7 shown. The computer device includes a processor, a memory, a communication interface, a display screen, and an input device connected through a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The communication interface of the computer device is used to communicate with an external terminal in a wired or wireless manner. The wireless manner can be achieved through WIFI, a mobile cellular network, NFC (Near Field Communication), or other technologies. When the computer program is executed by the processor, it implements a method for training a user preference prediction model and a method for predicting user preferences. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen. The input device of the computer device can be a touch layer covering the display screen, or a button, a trackball, or a touchpad provided on the housing of the computer device, or an external keyboard, touchpad, or mouse, etc.

[0143] Those skilled in the art can understand that Figure 7 the structure shown in

[0144] is only a block diagram of a part of the structure related to the solution of this application, and does not constitute a limitation on the computer device to which the solution of this application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0145] In one embodiment, a computer device is provided, including a memory and a processor. A computer program is stored in the memory, and when the processor executes the computer program, it implements the steps of the method described in any one of the first aspect or the second aspect.

[0146] In one embodiment, a computer program product is provided, including a computer program which, when executed by a processor, implements the steps of the method described in any one of the first aspect or the second aspect.

[0147] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data for analysis, stored data, displayed data, etc.) involved in this application are all information and data that have been authorized by the user or fully authorized by all parties.

[0148] Those of ordinary skill in the art can understand that all or part of the processes of implementing the methods in the above embodiments can be completed by instructing relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above methods. Among them, any reference to a memory, database, or other medium used in the embodiments provided in this application can include at least one of non-volatile and volatile memories. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetoresistive random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc. The databases involved in the embodiments provided in this application can include at least one of relational databases and non-relational databases. Non-relational databases can include distributed databases based on blockchain, etc., and are not limited thereto. The processors involved in the embodiments provided in this application can be general-purpose processors, central processors, graphics processors, digital signal processors, programmable logic devices, data processing logics based on quantum computing, etc., and are not limited thereto.

[0149] The technical features of the above embodiments can be combined arbitrarily. For the sake of concise description, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered as the scope recorded in this specification.

[0150] The above-described embodiments merely represent several implementation manners of the present application. The description is relatively specific and detailed, but it should not be construed as a limitation on the scope of the patent of the present application. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present application, several modifications and improvements can still be made, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the appended claims.

Claims

1. A training method for a user preference prediction model, characterized in that The method includes: Obtaining historical browsing sample information of a user; each piece of the historical browsing sample information includes historical click-browsing sample information of each user and historical preference-browsing sample information of the user; Performing dilated convolution processing on each piece of the historical click-browsing sample information through a convolutional layer of an initial user preference prediction model, and extracting click feature information of the historical click-browsing sample information after each dilated convolution processing through an extraction layer of the initial user preference prediction model; wherein, performing dilated convolution processing on each piece of the historical click-browsing sample information through the convolutional layer of the initial user preference prediction model, and extracting click feature information of the historical click-browsing sample information that has undergone dilated convolution processing through the extraction layer; in the case where the number of dilated convolution times has not reached a preset number of dilated convolution times, inputting the historical click-browsing sample information that has undergone dilated convolution processing into the convolutional layer, and returning to execute the step of performing dilated convolution processing on each piece of the historical click-browsing sample information through the convolutional layer of the initial user preference prediction model, and extracting click feature information of the historical click-browsing sample information that has undergone dilated convolution processing through the extraction layer, until the number of dilated convolution times reaches the preset number of dilated convolution times, and outputting the click feature information extracted each time; Predicting predicted preference information of the user according to each piece of the extracted click feature information through a prediction layer of the initial user preference prediction model, and training the initial user preference prediction model according to the preference-browsing sample information of the user and the predicted preference information of the user to obtain a user preference prediction model.

2. The method according to claim 1, wherein The training the initial user preference prediction model according to the preference-browsing sample information of the user and the predicted preference information of the user to obtain a user preference prediction model includes: Calculating the similarity between the predicted preference information and the preference-browsing sample information through a loss function; In the case where the similarity between the predicted preference information and the preference-browsing sample information is less than a similarity threshold, updating weight parameters of the prediction layer and extraction parameters of the extraction layer, and returning to execute the step of performing dilated convolution processing on each piece of the historical click-browsing sample information through the convolutional layer of the initial user preference prediction model until the similarity between the predicted preference information and the preference-browsing sample information is greater than the similarity threshold; Taking the initial user preference prediction model corresponding to the predicted preference information greater than the similarity threshold as the user preference prediction model.

3. The method according to claim 1, characterized in that, Before performing dilated convolution processing on each piece of the historical click-browsing sample information through the convolutional layer of the initial user preference prediction model, and extracting click feature information of the historical click-browsing sample information after each dilated convolution processing through the extraction layer of the initial user preference prediction model, it further includes: Extracting click feature information of each piece of the historical click-browsing sample information through the extraction layer of the initial user preference prediction model.

4. The method according to claim 1, wherein The click-browse sample information includes the sample click-browse rate. Predicting the predicted preference information of the user according to each extracted click feature information through the prediction layer of the initial user preference prediction model includes: Calculating the sample weight value corresponding to each click-browse sample information according to the sample click-browse rate of each click-browse sample information, and using the weight value corresponding to each click-browse sample information as the weight value of each click feature information extracted through the click-browse sample information in the prediction layer of the initial user preference prediction model; Performing weighted processing on each click feature information according to the weight values of each click feature information to obtain each target click feature information; According to each target click feature information, through the prediction layer of the initial user preference prediction model, screening the target historical browse sample information corresponding to each target click feature information in each historical browse sample information, and using the target historical browse sample information as the predicted preference information of the user.

5. A user preference prediction method, characterized in that, The method includes: Obtaining the historical click-browse information of the target user and the candidate information to be browsed by the target user; Performing dilated convolution processing on the historical click-browse information through the convolutional layer of the user preference prediction model, and extracting the click feature information of the click-browse information after each dilated convolution processing through the extraction layer of the user preference prediction model; Predicting the target preference browse information of the target user in each candidate information according to each click feature information through the prediction layer of the user preference prediction model; Wherein, the user preference prediction model is trained by the training method of the user preference prediction model according to any one of claims 1 to 4.

6. A training device for a user preference prediction model, characterized in that, The device includes: An acquisition module, configured to acquire the user's multiple historical browse sample information; each historical browse sample information includes each user's historical click-browse sample information and the user's historical preference browse sample information; A first feature extraction module, configured to perform dilated convolution processing on each historical click-browse sample information through the convolutional layer of the initial user preference prediction model, and extract the click feature information of the click-browse sample information after each dilated convolution processing through the extraction layer of the initial user preference prediction model; A training module, configured to predict the predicted preference information of the user according to each extracted click feature information through the prediction layer of the initial user preference prediction model, and train the initial user preference prediction model according to the user's preference browse sample information and the user's predicted preference information to obtain a user preference prediction model; The first feature extraction module is further configured to: perform dilated convolution processing on each of the historical click-browse sample information through the convolutional layer of the initial user preference prediction model, and extract click feature information of the dilated-convolved historical click-browse sample information through the extraction layer; in the case where the number of dilated convolutions has not reached the preset number of dilated convolutions, input the dilated-convolved historical click-browse sample information into the convolutional layer, and return to execute the step of performing dilated convolution processing on each of the historical click-browse sample information through the convolutional layer of the initial user preference prediction model, and extracting click feature information of the dilated-convolved historical click-browse sample information through the extraction layer, until the number of dilated convolutions reaches the preset number of dilated convolutions, and output the click feature information extracted each time.

7. A user preference prediction device, characterized in that, The device includes: A target information acquisition module, configured to acquire the historical click-browse information of a target user and the candidate information to be browsed by the target user; A feature extraction module, configured to perform dilated convolution processing on the historical click-browse information through the convolutional layer of a user preference prediction model, and extract click feature information of the click-browse information after each dilated convolution processing through the extraction layer of the user preference prediction model; A prediction module, configured to predict the target preference browse information of the target user among each of the candidate information through the prediction layer of the user preference prediction model according to each of the click feature information; Wherein, the user preference prediction model is trained by the training method of the user preference prediction model according to any one of claims 1 to 4.

8. A computer device, comprising a memory and a processor, the memory storing a computer program, characterized in that, When the processor executes the computer program, the steps of the method according to any one of claims 1 to 4 or 5 are implemented.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, the steps of the method according to any one of claims 1 to 4 or 5 are implemented.

10. A computer program product comprising a computer program, characterized in that, When the computer program is executed by the processor, the steps of the method according to any one of claims 1 to 4 or 5 are implemented.

Images