A model training method and apparatus for determining recommended works

Abstract

This invention provides a method and apparatus for training a recommended works determination model, relating to the field of data processing technology. The method includes: acquiring historical behavior data of sample users; extracting sample content feature information and sample cover image feature information to obtain sample user feature information; generating sample feature information based on the sample content feature information, sample cover image feature information, and sample user feature information; inputting the sample feature information into a recommended works determination model to obtain an evaluation score; calculating the sample loss of the recommended works determination model based on the evaluation score and the historical behavior data of the sample users; adjusting the parameters of the recommended works determination model based on the sample loss; and returning to the step of inputting the sample feature information into the recommended works determination model until a preset training convergence condition is reached, resulting in a trained recommended works determination model. The solution provided by this invention can be used to train a model capable of recommending works to users.

Description

Technical Field

[0001] The embodiments of the present invention relate to the field of data processing technology, and in particular to a recommended work determination model training, a recommended work determination method and apparatus. Background Technology

[0002] Currently, with the development of internet technology, audiovisual entertainment has become increasingly important, and users hope to find their favorite entertainment content on various online platforms. In this context, each platform is striving to improve the quality of its recommendations to better meet user needs and attract and retain users. Therefore, recommending content to users is a key focus for all platforms. Summary of the Invention

[0003] The purpose of this invention is to provide a model training, method, and apparatus for determining recommended works, so as to train a model capable of recommending works to users. The specific technical solution is as follows:

[0004] In a first aspect, embodiments of the present invention provide a method for training a recommended works determination model, the method comprising:

[0005] Obtain historical behavior data of sample users;

[0006] Extract the sample content feature information of at least one sample work and the sample cover image feature information of the sample work from the historical behavior data, and obtain the sample user feature information of the sample user;

[0007] Based on the sample content feature information, sample cover image feature information, and sample user feature information, sample feature information is generated;

[0008] The sample feature information is input into the recommended work determination model to obtain the evaluation score of the sample work;

[0009] Based on the evaluation scores and the historical behavior data of the sample users, the sample loss of the recommended work determination model is calculated;

[0010] Based on the sample loss, the parameters of the recommended work determination model are adjusted, and the step of inputting the sample feature information into the recommended work determination model is returned to be executed until the preset training convergence condition is reached, and the trained recommended work determination model is obtained.

[0011] In one embodiment of the present invention, when the sample work has multiple cover images, the sample cover image of the sample work is obtained in the following manner:

[0012] Based on the weight factors of each cover image, a target number of sample cover images are selected. Cover images with larger weight factors have higher historical click rates and are more likely to be selected.

[0013] In one embodiment of the present invention, after selecting the target number of sample cover images, the method further includes:

[0014] If the selected target number of sample cover images does not include the cover image shown to the sample user, the selected sample cover image is replaced with the shown cover image.

[0015] In one embodiment of the present invention, after selecting the target number of sample cover images, the method further includes:

[0016] If the selected target number of sample cover images does not include the cover image shown to the sample user, the shown cover image will be added to the sample cover images.

[0017] In one embodiment of the present invention, the recommended work determination model includes multiple sub-networks, and the feature information of the sample cover image is processed in the following manner:

[0018] The feature information of the sample cover image is processed by multiple sub-networks, and the processing results output by each sub-network are obtained.

[0019] The processing results are combined and calculated based on the weights of each sub-network.

[0020] Secondly, embodiments of the present invention provide a method for determining recommended works, the method comprising:

[0021] Obtain the content feature information of the work to be recommended and the cover image feature information of the cover image of the work to be recommended;

[0022] Obtain target user characteristic information;

[0023] Based on the content feature information, cover image feature information, and target user feature information, target feature information is generated;

[0024] The target feature information is input into a pre-trained recommended works determination model, which is trained according to the method steps described in any one of the first aspects;

[0025] The output of the recommended work determination model is obtained to determine whether the work to be recommended is a recommended work.

[0026] In one embodiment of the present invention, when the work to be recommended has at least two cover images, the step of generating target feature information based on the content feature information, cover image feature information, and target user feature information includes:

[0027] For each of the at least two cover images, target feature information corresponding to the cover image is generated based on the content feature information, the cover image feature information of the cover image, and the target user feature information.

[0028] Thirdly, embodiments of the present invention provide a training device for a recommended works determination model, the device comprising:

[0029] The first acquisition module is used to acquire historical behavior data of sample users;

[0030] The extraction module is used to extract the sample content feature information of at least one sample work and the sample cover image feature information of the sample work from the historical behavior data, and to obtain the sample user feature information of the sample user.

[0031] The first generation module is used to generate sample feature information based on the sample content feature information, sample cover image feature information and sample user feature information;

[0032] The input and output module is used to input the sample feature information into the recommendation work determination model to obtain the evaluation score of the sample work;

[0033] The loss calculation module is used to calculate the sample loss of the recommended work determination model based on the evaluation score and the historical behavior data of the sample users.

[0034] The adjustment module is used to adjust the parameters of the recommended work determination model based on the sample loss, and then return to the step of inputting the sample feature information into the recommended work determination model until the preset training convergence condition is reached, so as to obtain the trained recommended work determination model.

[0035] Fourthly, embodiments of the present invention provide a recommended works determination device, the device comprising:

[0036] The second acquisition module is used to acquire the content feature information of the work to be recommended and the cover image feature information of the cover image of the work to be recommended;

[0037] The third acquisition module is used to acquire target user characteristic information of the target user;

[0038] The second generation module is used to generate target feature information based on the content feature information, cover image feature information and target user feature information;

[0039] The input module is used to input the target feature information into a pre-trained recommended works determination model, which is trained according to the method steps described in any one of the first aspects;

[0040] The determination module is used to obtain the output result of the recommended work determination model and determine whether the work to be recommended is a recommended work.

[0041] Fifthly, embodiments of the present invention provide an electronic device, including a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus;

[0042] Memory, used to store computer programs;

[0043] A processor, when executing a program stored in memory, implements the method steps of any one of the first or second aspects.

[0044] In a sixth aspect, embodiments of the present invention provide a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the method steps of any one of the first or second aspects.

[0045] Beneficial effects of the embodiments of the present invention:

[0046] This invention provides a method for training a recommended work determination model, which involves: acquiring historical behavior data of sample users; extracting sample content feature information and sample cover image feature information of at least one sample work from the historical behavior data, and acquiring sample user feature information of the sample users; generating sample feature information based on the sample content feature information, sample cover image feature information, and sample user feature information; inputting the sample feature information into the recommended work determination model to obtain an evaluation score for the sample work; calculating the sample loss of the recommended work determination model based on the evaluation score and the historical behavior data of the sample users; adjusting the parameters of the recommended work determination model based on the sample loss; and returning to the step of inputting the sample feature information into the recommended work determination model until a preset training convergence condition is reached to obtain a trained recommended work determination model.

[0047] As can be seen from the above, in the solution provided by the embodiments of the present invention, the parameters of the recommended work determination model are continuously adjusted through each training iteration, ultimately resulting in a well-trained recommended work determination model. In the recommended work determination model, by processing sample feature information, the content score and cover image score of the sample work are obtained in one step. Then, based on the content score and cover image score, an evaluation score is obtained for the sample work. Subsequently, sample works whose evaluation scores meet preset conditions can be determined as recommended works. The sample feature information includes the content feature information and cover image feature information of the sample work. In the process of calculating the evaluation score of the sample work, both the content and cover image of the sample work are considered simultaneously. Compared with considering only the content or cover image of the sample work, this provides a more comprehensive evaluation of the value of the sample work, improving the recommendation quality of the subsequently determined recommended works.

[0048] Furthermore, by comprehensively considering the characteristics of sample users, the content of sample works, and cover images, the recommended works determined based on evaluation scores are more applicable to real-world scenarios and can more accurately and realistically identify the works preferred by sample users. Analyzing the preferences of sample users can provide auxiliary decision-making analysis for the content platform, offering a basis for improving related products and services. Attached Figure Description

[0049] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other embodiments can be obtained based on these drawings.

[0050] Figure 1 A flowchart illustrating a recommended works determination model training method provided in an embodiment of the present invention;

[0051] Figure 2 A flowchart illustrating a method for determining recommended works provided in an embodiment of the present invention;

[0052] Figure 3 This is a schematic diagram of the structure of a recommended work determination model training device provided in an embodiment of the present invention;

[0053] Figure 4 This is a schematic diagram of a recommended work determination device provided in an embodiment of the present invention;

[0054] Figure 5 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention;

[0055] Figure 6 This is a schematic diagram of another electronic device provided in an embodiment of the present invention. Detailed Implementation

[0056] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art based on the present invention are within the scope of protection of the present invention.

[0057] In related technologies, online platforms recommend works to users primarily through two stages. The first stage involves recommending content, which can be done by sorting works using personalized recommendation methods such as recall ranking, and then selecting the works to recommend to the user. The second stage involves selecting a suitable cover image for each recommended work. Since users have different preferences for different cover images of the same work, the technology recommends cover images based on user click feedback in the second stage.

[0058] A cover image that users like is more likely to attract users to click on the work corresponding to that cover image. However, the technology that recommends works to users through the above two stages fails to take into account the impact of both the work content and the cover image on the recommendation effect, which leads to limited recommendation effectiveness. The technology that recommends works through the above two stages has certain limitations.

[0059] To address the aforementioned issues, embodiments of the present invention provide a recommended works determination model training method and apparatus.

[0060] First, the recommended works determination model training method provided by the embodiments of the present invention will be described.

[0061] See Figure 1 This is a flowchart illustrating a method for training a recommended works determination model according to an embodiment of the present invention. This method can be applied to electronic devices with computing capabilities; for example, it is applied to a server. The method includes the following steps S101 to S106.

[0062] Step S101: Obtain historical behavior data of sample users.

[0063] Specifically, obtaining historical behavioral data of sample users refers to acquiring works that have been recommended to the sample users and data related to the sample users' viewing behavior of these recommended works, such as whether the sample users clicked on the works. These works include content and cover information, and can be audio / video materials, publications (magazines, novels), etc.

[0064] Step S102: Extract the sample content feature information of at least one sample work and the sample cover image feature information of the sample cover image of the above-mentioned sample work from the above-mentioned historical behavior data, and obtain the sample user feature information of the above-mentioned sample user.

[0065] Specifically, the sample works mentioned above may have been viewed by the aforementioned sample users, or they may not have been viewed by them. Sample works viewed by the aforementioned sample users are considered positive sample works, and sample works not viewed by the aforementioned sample users are considered negative sample works.

[0066] For the aforementioned negative sample works, the sample content feature information is a brief description of the content of the negative sample work. For example, the sample content feature information may include the duration, type, and tags of the negative sample work. Here, the type refers to different categories or combinations of works such as comedy, period drama, and suspense. For example, the type of the aforementioned negative sample work is comedy, or the type of the aforementioned negative sample work is period comedy. The tags refer to the keywords of the aforementioned negative sample work. For example, the tags of the aforementioned negative sample work are: anime, action, and youth.

[0067] This invention not only acquires the sample content feature information of the aforementioned negative sample works, but also acquires the sample cover image feature information of the sample cover images of the aforementioned negative sample works. The sample cover image feature information may include the color distribution, image theme, number of people, etc., of the sample cover image. Specifically, the color distribution may refer to the proportion of each color in the image and the position of each color within the image; the image theme refers to the image style of the sample cover image; and the number of people refers to the number of figures depicted in the sample cover image.

[0068] Compared with the sample content feature information of the negative sample works, the sample content feature information of the positive sample works also includes the ratings of the sample users for the content of the positive sample works; compared with the sample cover image feature information of the negative sample works, the sample cover image feature information of the positive sample works also includes the quality ratings of the sample cover images of the positive sample works by the sample users.

[0069] The aforementioned sample user characteristics may include the sample users' age, gender, interests, etc.

[0070] Step S103: Generate sample feature information based on the above sample content feature information, sample cover image feature information and sample user feature information.

[0071] After obtaining the above-mentioned sample content feature information, sample cover image feature information and sample user feature information, these three types of feature information are fused to generate the above-mentioned sample feature information, which includes the above-mentioned sample content feature information, sample cover image feature information and sample user feature information.

[0072] Specifically, the sample content feature information, sample cover image feature information, and sample user feature information obtained above can each be represented in the form of a vector. The vectors are combined into a single sample vector, which is the sample feature information mentioned above.

[0073] Step S104: Input the above sample feature information into the recommended works determination model to obtain the evaluation score of the above sample works.

[0074] Specifically, in the process of the aforementioned recommended work determination model processing the sample feature information, it calculates the sample content score of the sample work based on the sample content feature information contained in the sample feature information, and calculates the sample cover image score based on the sample cover image feature information contained in the sample feature information. The sample content score represents the degree to which the sample user likes the sample work's content, and the sample cover image score represents the degree to which the sample user likes the sample cover image. For example, a higher score indicates a higher degree of liking.

[0075] In the aforementioned model for determining recommended works, an evaluation score for each work is obtained by comprehensively considering both the sample content score and the sample cover image score. Specifically, the evaluation score for each work is calculated by weighting the sample content score and the sample cover image score according to their respective weights.

[0076] In one embodiment of the present invention, the aforementioned recommended work determination model includes a DNN (Deep Neural Network), an ESMM (Entire Space Multi-Task Model), an MMOE (Multi-gate Mixture-of-Experts), and a scoring evaluation unit. The DNN is used to compress the input information, the ESMM is used to predict the user's liking for the work content based on the compressed information to obtain the work's content score, the MMOE is used to predict the user's liking for the cover image based on the compressed information to obtain the cover image score, and the scoring evaluation unit is used to calculate the work's evaluation score based on the content score and the cover image score.

[0077] Since the amount of sample feature information can be large, and when the sample feature information is represented by sample vectors, the dimensionality of the sample vectors may be high. In this case, the sample vectors can be compressed first to obtain compressed sample vectors, thereby reducing the vector dimensionality and facilitating subsequent processing. Specifically, the sample vectors can be input into a DNN, and after being processed by the hidden layers of the DNN, the compressed sample vectors are obtained.

[0078] Furthermore, to calculate the content score of a sample work, a compressed sample vector representing the sample's feature information can be input into ESMM. ESMM has two objective functions: CTR (Click-through rate) and CVR (Conversion Rate). After ESMM processes the input compressed sample vector to obtain CTR and CVR, the content score of the sample work is obtained based on the obtained CTR and CVR. Specifically, the product of CTR and CVR can be used as the content score of the sample work.

[0079] For example, if the sample work is a video, then the sample work is referred to as a sample video. CTR can refer to the ratio of the number of clicks on the sample video to the number of impressions of the sample video. If the sample video is a free video, CVR can refer to the ratio of the number of successful user registrations after clicking on the sample video to the number of clicks on the sample video. If the sample video is a paid video, CVR can refer to the ratio of the number of successful payments after clicking on the sample video to the number of clicks on the sample video.

[0080] To calculate the score of the sample cover image of a sample work, specifically, the compressed sample vector representing the sample feature information can be input into MMOE, and the sample cover image score of the sample work can be obtained after processing by MMOE.

[0081] Step S105: Based on the above evaluation scores and the historical behavior data of the above sample users, calculate the sample loss of the above recommended works determination model.

[0082] After obtaining the evaluation scores of the sample works, the sample loss of the recommendation work determination model is calculated by combining the historical behavior data of the sample users, specifically the actual CTR and CVR of the sample works. Theoretically, the higher the actual CTR and CVR of the sample works, the higher the evaluation score of the sample works should be.

[0083] Step S106: Adjust the parameters of the above-mentioned recommended work determination model based on the above-mentioned sample loss, and return to the above-mentioned step of inputting the above-mentioned sample feature information into the recommended work determination model until the preset training convergence condition is reached, and the trained recommended work determination model is obtained.

[0084] The loss function is optimized by adjusting the parameters. After each parameter adjustment, step S104 and subsequent steps are executed until the preset training convergence condition is reached.

[0085] Specifically, the sample works input into the above-mentioned recommended works determination model are divided into a training set and a validation set. Early stopping can be used to monitor the output results of the above-mentioned recommended works determination model on the validation set to prevent overfitting. At the same time, techniques such as learning rate decay, regularization, and dropout can also be used to optimize the above-mentioned recommended works determination model.

[0086] Furthermore, by continuously collecting data on works recommended to sample users, new sample feature information is obtained. This data is then used to periodically update the training model for determining recommended works. As needed, some weight coefficients in the model can be dynamically adjusted, and the model's network structure improved to continuously optimize the model. Additionally, information from other modalities, such as audio and comments from works recommended to sample users, can be incorporated to allow the model to more comprehensively evaluate the value of the works. The aforementioned model for determining recommended works can be developed using Python and implemented using relevant machine learning libraries, such as TensorFlow.

[0087] As can be seen from the above, in the solution provided by the embodiments of the present invention, the parameters of the recommended work determination model are continuously adjusted through each training iteration, ultimately resulting in a well-trained recommended work determination model. In the recommended work determination model, by processing sample feature information, the content score and cover image score of the sample work are obtained in one step. Then, based on the content score and cover image score, an evaluation score is obtained for the sample work. Subsequently, sample works whose evaluation scores meet preset conditions can be determined as recommended works. The sample feature information includes the content feature information and cover image feature information of the sample work. In the process of calculating the evaluation score of the sample work, both the content and cover image of the sample work are considered simultaneously. Compared with considering only the content or cover image of the sample work, this provides a more comprehensive evaluation of the value of the sample work, improving the recommendation quality of the subsequently determined recommended works.

[0088] Furthermore, by comprehensively considering the characteristics of sample users, the content of sample works, and cover images, the recommended works determined based on evaluation scores are more applicable to real-world scenarios and can more accurately and realistically identify the works preferred by sample users. Analyzing the preferences of sample users can provide auxiliary decision-making analysis for the content platform, offering a basis for improving related products and services.

[0089] If the number of cover images for a sample work is too large, processing all cover images to generate corresponding sample feature information may affect the efficiency of the recommendation work determination model. In one embodiment of the present invention, the sample work has multiple cover images, and the sample cover images are obtained through the following step A.

[0090] Step A: Select the target number of sample cover images based on the weighting factors of each cover image.

[0091] Among them, the larger the weight factor, the higher the historical click rate of the cover image, and the greater the probability of it being selected.

[0092] For each cover image in the sample works, each cover image has the potential to be selected as one of the target number of sample cover images, but cover images with higher weighting factors have a higher probability of being selected. Furthermore, the target number can be set according to requirements.

[0093] After selecting the target number of sample cover images, the cover image feature information of the target number of sample cover images is determined as the sample cover image feature information of the sample works.

[0094] Specifically, the MAB (Multi-Armed Bandit) algorithm can be used to select a target number of sample cover images. The cover image features and historical click-through rates of each sample cover image are input into the MAB algorithm. The MAB algorithm assigns a weight factor to each cover image based on its historical click-through rate and selects the target number of sample cover images based on this weight factor. Cover images with larger weight factors have a higher probability of being selected, thus discovering new potential high-quality sample cover images and avoiding over-reliance on historically high-performing sample cover images.

[0095] As can be seen from the above, in the solution provided by the embodiments of the present invention, when there are too many cover images of sample works, a target number of sample cover images are selected to reduce the number of cover images that need to be processed, thereby ensuring the running efficiency of the recommended work determination model. Furthermore, based on the weight factors of each cover image, a target number of sample cover images are selected. Cover images with larger weight factors have a higher probability of being selected, while cover images with smaller weight factors also have a probability of being selected, although this probability is small, it is not zero. Therefore, in the process of selecting a target number of sample cover images, each cover image has the possibility of being selected, ensuring the diversity of selection results.

[0096] In one embodiment of the present invention, after selecting a target number of sample cover images through step A above, the following step B is also included.

[0097] Step B: If the selected target number of sample cover images does not include the cover image shown to the above sample users, replace the selected sample cover image with the shown cover image.

[0098] After replacing the selected sample cover image with the displayed cover image, a new target number of sample cover images are obtained, which contain the displayed cover image. Then, the new target number of sample cover images are processed to obtain the sample cover image feature information.

[0099] Specifically, if it is found that the target number of sample cover images selected using the MAB algorithm do not include the cover image shown to the aforementioned sample users, then the selected sample cover images are replaced with the displayed cover images, and the parameters of the MAB algorithm are adjusted based on the cover image feature information of the new target number of sample cover images.

[0100] In another embodiment of the present invention, after selecting the target number of sample cover images through step A above, the following step C is also included.

[0101] Step C: If the selected target number of sample cover images does not include the cover image shown to the above sample users, add the shown cover image to the sample cover images.

[0102] Specifically, the displayed cover image will be added to the sample cover image. In other words, the target number of sample cover images selected in step A and the cover image shown above will be processed to obtain the feature information of the sample cover image.

[0103] As can be seen from the above, in the solution provided by the embodiments of the present invention, the cover image feature information processed by the recommended work determination model includes the cover image feature information of the cover image shown to the sample users. For a cover image shown to the sample users, if the sample users have viewed the work corresponding to that cover image, it indicates that the cover image is relatively popular with the sample users; if the sample users have not viewed the work corresponding to that cover image, it indicates that the cover image is less popular with the sample users. In both cases, it indicates that the cover image is representative and can reflect the sample users' preference for similar cover images. Therefore, by considering the cover image shown to the sample users, the recommended work determination model can make the evaluation score of the obtained sample work more consistent with the sample users' preference for that sample work.

[0104] In one embodiment of the present invention, the above-mentioned recommended works determination model contains multiple sub-networks, and the feature information of the sample cover image is processed through the following steps D and E.

[0105] Step D: Process the feature information of the sample cover image through multiple sub-networks respectively, and obtain the processing results output by each sub-network.

[0106] The parameters of different subnetworks in the above-mentioned subnetworks are different, and the focus of processing the sample cover image feature information is different. For example, one subnetwork focuses on the sample cover image color information in the sample cover image feature information, while another subnetwork focuses on the sample cover image type information in the sample cover image feature information.

[0107] After each subnetwork processes the feature information of the sample cover image, the output result can be a score.

[0108] Step E: Combine and calculate the processing results based on the weights of each sub-network.

[0109] Each sub-network outputs its corresponding weight, and the processing results of each sub-network are combined based on these weights. For example, if each sub-network outputs its own score, then the scores can be weighted and averaged, and this result can be used as the processing result for the feature information of the sample cover image.

[0110] Specifically, the aforementioned sub-network can be the expert network in MMOE. In the process of processing the feature information of the sample cover image, MMOE learns the weights of each expert network in MMOE through Gating Network. Multiple expert networks predict the sample cover image score of the sample cover image respectively. The sample cover image scores predicted by multiple expert networks are summed according to their weights to obtain the final sample cover image score of the sample work.

[0111] As can be seen from the above, in the solution provided by the embodiments of the present invention, the sample cover image feature information is processed by multiple sub-networks respectively, and the processing results of each sub-network are merged and calculated according to weights, so as to obtain the processing result of the sample cover image feature information.

[0112] See Figure 2 The above is a flowchart illustrating a method for determining recommended works according to an embodiment of the present invention. The method is applied to an electronic device with computing capabilities and includes the following steps S201 to S205.

[0113] Step S201: Obtain the content feature information of the work to be recommended and the cover image feature information of the cover image of the work to be recommended.

[0114] For the aforementioned pre-prepared works to be recommended, obtain the content feature information and the cover image feature information of the works to be recommended. Specifically, the content feature information and cover image feature information of the works to be recommended are the same as the sample content feature information and sample cover image feature information of the negative samples described in step S102.

[0115] Step S202: Obtain the target user characteristic information of the target user.

[0116] In order to recommend works to the target user, it is necessary to obtain the user characteristic information of the target user, namely the target user characteristic information mentioned above. Specifically, the target user characteristic information mentioned above is the same as the sample user characteristic information described in step S102, which can include the target user's age, gender, interests, etc.

[0117] Step S203: Based on the above content feature information, cover image feature information and target user feature information, generate target feature information.

[0118] Specifically, the method for generating the above target feature information can be referred to the description in step S103.

[0119] Step S204: Input the above target feature information into the pre-trained recommendation work determination model.

[0120] The aforementioned recommended work determination model was trained according to the method steps outlined in the training method for any of the aforementioned recommended work determination models. A description of the recommended work determination model can be found in the relevant content of the training method for the aforementioned recommended work determination model.

[0121] Step S205: Obtain the output of the above-mentioned recommended work determination model, and determine whether the above-mentioned work to be recommended is a recommended work.

[0122] Specifically, the processing of the target feature information by the above-mentioned recommended works determination model can be found in the description of sample feature information processing in the training method of the above-mentioned recommended works determination model.

[0123] After the recommended work determination model processes the target feature information and outputs an evaluation score for the works to be recommended, it needs to consider whether the evaluation score meets a preset condition to determine whether to determine whether to identify the works to be recommended as recommended works. Specifically, the preset condition can be a preset score; the works to be recommended are only identified as recommended works if the evaluation score is not less than the preset score. Alternatively, if multiple works to be recommended have their own evaluation scores, the works ranked in the top preset percentage can be identified as recommended works.

[0124] In addition, the recommendation results can be adjusted based on user feedback and behavior regarding the selected recommended works, such as clicks, views, and ratings.

[0125] In one example, if a user clicks on or views a target recommended work, in the subsequent process of determining recommended works, works of the same type as the target recommended work will be identified as recommended works.

[0126] In another example, if a user's rating of the identified target recommended work reaches a preset score, then in the subsequent process of determining recommended works, works of the same type as the target recommended work will be identified as recommended works.

[0127] As can be seen from the above, in the solution provided by the embodiments of the present invention, the above-mentioned target feature information is input into a pre-trained recommendation work determination model to obtain the evaluation score of the work to be recommended, and then the evaluation score can be used to determine whether the work to be recommended is recommended to the target user.

[0128] Furthermore, since the aforementioned target feature information includes the content and cover image features of the works to be recommended, the evaluation score obtained using the aforementioned work recommendation determination model considers both the content and cover image of the works to be recommended. Compared to considering only the content or cover image, this approach provides a more comprehensive assessment of the value of the works to be recommended, thus improving the quality of the final recommended works. By comprehensively considering the features of the target user, the content of the works to be recommended, and the cover image, the final recommended works are more applicable to real-world scenarios, more accurately and realistically identifying the works that target users like. This enhances the viewing experience for target users, increases the time they spend on the platform and their retention rate, and ultimately improves the click-through rate of advertisements on the platform. Different recommended works are determined for different target users, enriching personalized recommendation strategies and allowing advertisers to segment target customer groups. By comprehensively considering the content and cover image of the works to be recommended, resources can be used more rationally, resource allocation priorities can be adjusted, and resource utilization efficiency can be improved.

[0129] In addition, by using the above-mentioned recommended works determination model to consider both the content and cover image of the works, the resources consumed in calculating the evaluation scores of the works and thus determining the recommended works can be saved; the model can be applied to different fields that consider both the content and cover image of the works, and has scalability.

[0130] Typically, a work may have multiple cover images. In the case where the work to be recommended has at least two cover images, in one embodiment of the present invention, step S203 can be implemented by step F.

[0131] Step F: For each of the at least two cover images, based on the above content feature information, the cover image feature information, and the target user feature information, generate the target feature information corresponding to the cover image.

[0132] For each cover image of the aforementioned works to be recommended, after obtaining the cover image feature information, the target feature information corresponding to the cover image can be generated according to the description in step S103 above. Then, the generated target feature information can be processed according to step S204 above, and the evaluation score of the work to be recommended corresponding to the cover image can be obtained according to step S205 above. Finally, the recommended works are determined based on the evaluation scores.

[0133] Furthermore, if a work to be recommended has at least two cover images, an evaluation score can be calculated for each cover image. This means that a single work to be recommended will receive multiple evaluation scores, the number of which is the same as the number of cover images. In this case, during the process of determining recommended works, for multiple works with the same content but different cover images, the one with the highest evaluation score can be retained, and the final recommendation can be made based on this score.

[0134] As can be seen from the above, the solution provided by the embodiments of the present invention can be applied to the case where there are at least two cover images of the work to be recommended. For each cover image, based on the content feature information of the work to be recommended, the cover image feature information of the cover image, and the target user feature information, the target feature information corresponding to the cover image is generated. Then, the target feature information is processed to obtain an evaluation score and determine the recommended work.

[0135] If there are too many cover images of works to be recommended, processing all of them may affect the efficiency of determining the recommended works. Therefore, a predetermined target number of cover images can be selected from all the cover images of the works to be recommended. The method for selecting the predetermined target number of cover images can be referred to the description of selecting the target number of sample cover images in the training method of the recommended work determination model above.

[0136] Specifically, the MAB algorithm from the aforementioned recommended works determination model training method, or other related algorithms, can be used to process all the cover images of the works to be recommended. The MAB algorithm assigns a weight factor to each cover image and selects a target number of cover images based on the weight factor. The cover image with the larger the weight factor has a higher probability of being selected.

[0137] Corresponding to the aforementioned method for training a recommended works determination model, this embodiment of the invention also provides a device for training a recommended works determination model.

[0138] See Figure 3 The above is a schematic diagram of a recommended works determination model training device provided in an embodiment of the present invention. The device includes:

[0139] The first acquisition module 301 is used to acquire historical behavior data of sample users.

[0140] The extraction module 302 is used to extract the sample content feature information of at least one sample work and the sample cover image feature information of the sample work from the above-mentioned historical behavior data, and to obtain the sample user feature information of the above-mentioned sample user.

[0141] The first generation module 303 is used to generate sample feature information based on the above-mentioned sample content feature information, sample cover image feature information and sample user feature information.

[0142] The input and output module 304 is used to input the above sample feature information into the recommended works determination model to obtain the evaluation score of the above sample works.

[0143] The loss calculation module 305 is used to calculate the sample loss of the recommended works determination model based on the above evaluation scores and the historical behavior data of the above sample users.

[0144] The adjustment module 306 is used to adjust the parameters of the recommended works determination model based on the above sample loss, and return to execute the above steps of inputting the above sample feature information into the recommended works determination model until the preset training convergence condition is reached, so as to obtain the trained recommended works determination model.

[0145] As can be seen from the above, in the solution provided by the embodiments of the present invention, the parameters of the recommended work determination model are continuously adjusted through each training iteration, ultimately resulting in a well-trained recommended work determination model. In the recommended work determination model, by processing sample feature information, the content score and cover image score of the sample work are obtained in one step. Then, based on the content score and cover image score, an evaluation score is obtained for the sample work. Subsequently, sample works whose evaluation scores meet preset conditions can be determined as recommended works. The sample feature information includes the content feature information and cover image feature information of the sample work. In the process of calculating the evaluation score of the sample work, both the content and cover image of the sample work are considered simultaneously. Compared with considering only the content or cover image of the sample work, this provides a more comprehensive evaluation of the value of the sample work, improving the recommendation quality of the subsequently determined recommended works.

[0146] Furthermore, by comprehensively considering the characteristics of sample users, the content of sample works, and cover images, the recommended works determined based on evaluation scores are more applicable to real-world scenarios and can more accurately and realistically identify the works preferred by sample users. Analyzing the preferences of sample users can provide auxiliary decision-making analysis for the content platform, offering a basis for improving related products and services.

[0147] In one embodiment of the present invention, the above-mentioned sample work has multiple cover images, and the above-mentioned device further includes:

[0148] The selection module 307 is used to select a target number of sample cover images based on the weight factors of each cover image. The cover image with the larger weight factor has a higher historical click rate and a higher probability of being selected.

[0149] As can be seen from the above, in the solution provided by the embodiments of the present invention, when there are too many cover images of sample works, a target number of sample cover images are selected to reduce the number of cover images that need to be processed, thereby ensuring the running efficiency of the recommended work determination model. Furthermore, based on the weight factors of each cover image, a target number of sample cover images are selected. Cover images with larger weight factors have a higher probability of being selected, while cover images with smaller weight factors also have a probability of being selected, although this probability is small, it is not zero. Therefore, in the process of selecting a target number of sample cover images, each cover image has the possibility of being selected, ensuring the diversity of selection results.

[0150] In one embodiment of the present invention, the above-described apparatus further includes:

[0151] Replacement module 308 is used to replace the selected sample cover image with the displayed cover image when the selected target number of sample cover images do not include the cover image shown to the aforementioned sample users.

[0152] In another embodiment of the present invention, the above-described apparatus further includes:

[0153] Add module 309 to add the displayed cover image to the sample cover image if the selected target number of sample cover images do not include the cover image shown to the above sample users.

[0154] As can be seen from the above, in the solution provided by the embodiments of the present invention, the cover image feature information processed by the recommended work determination model includes the cover image feature information of the cover image shown to the sample users. For a cover image shown to the sample users, if the sample users have viewed the work corresponding to that cover image, it indicates that the cover image is relatively popular with the sample users; if the sample users have not viewed the work corresponding to that cover image, it indicates that the cover image is less popular with the sample users. In both cases, it indicates that the cover image is representative and can reflect the sample users' preference for similar cover images. Therefore, by considering the cover image shown to the sample users, the recommended work determination model can make the evaluation score of the obtained sample work more consistent with the sample users' preference for that sample work.

[0155] In one embodiment of the present invention, the above-mentioned recommended works determination model includes multiple sub-networks, and the above-mentioned input and output module 304 includes:

[0156] The network processing submodule 3041 is used to process the feature information of the sample cover image through multiple subnetworks and obtain the processing results output by each subnetwork.

[0157] The merge calculation submodule 3042 is used to merge and calculate the processing results based on the weights of each sub-network.

[0158] As can be seen from the above, in the solution provided by the embodiments of the present invention, the sample cover image feature information is processed by multiple sub-networks respectively, and the processing results of each sub-network are merged and calculated according to weights, so as to obtain the processing result of the sample cover image feature information.

[0159] Corresponding to the aforementioned method for determining recommended works, this embodiment of the invention also provides a device for determining recommended works.

[0160] See Figure 4 The diagram below illustrates the structure of a recommended work determination device according to an embodiment of the present invention. The device includes:

[0161] The second acquisition module 401 is used to acquire the content feature information of the work to be recommended and the cover image feature information of the cover image of the work to be recommended.

[0162] The third acquisition module 402 is used to acquire the target user feature information of the target user.

[0163] The second generation module 403 is used to generate target feature information based on the above-mentioned content feature information, cover image feature information and target user feature information.

[0164] The input module 404 is used to input the above target feature information into a pre-trained recommended work determination model, which is trained according to the method steps in any of the above recommended work determination model training methods.

[0165] The determination module 405 is used to obtain the output results of the above-mentioned recommended work determination model and determine whether the above-mentioned work to be recommended is a recommended work.

[0166] As can be seen from the above, in the solution provided by the embodiments of the present invention, the above-mentioned target feature information is input into a pre-trained recommendation work determination model to obtain the evaluation score of the work to be recommended, and then the evaluation score can be used to determine whether the work to be recommended is recommended to the target user.

[0167] Furthermore, since the aforementioned target feature information includes the content and cover image features of the works to be recommended, the evaluation score obtained using the aforementioned work recommendation determination model considers both the content and cover image of the works to be recommended. Compared to considering only the content or cover image, this approach provides a more comprehensive assessment of the value of the works to be recommended, thus improving the quality of the final recommended works. By comprehensively considering the features of the target user, the content of the works to be recommended, and the cover image, the final recommended works are more applicable to real-world scenarios, more accurately and realistically identifying the works that target users like. This enhances the viewing experience for target users, increases the time they spend on the platform and their retention rate, and ultimately improves the click-through rate of advertisements on the platform. Different recommended works are determined for different target users, enriching personalized recommendation strategies and allowing advertisers to segment target customer groups. By comprehensively considering the content and cover image of the works to be recommended, resources can be used more rationally, resource allocation priorities can be adjusted, and resource utilization efficiency can be improved.

[0168] In addition, by using the above-mentioned recommended works determination model to consider both the content and cover image of the works, the resources consumed in calculating the evaluation scores of the works and thus determining the recommended works can be saved; the model can be applied to different fields that consider both the content and cover image of the works, and has scalability.

[0169] In one embodiment of the present invention, the aforementioned work to be recommended has at least two cover images, and the second generation module 403 is specifically used for:

[0170] For each of the at least two cover images, target feature information corresponding to that cover image is generated based on the aforementioned content feature information, the cover image feature information of that cover image, and the target user feature information.

[0171] As can be seen from the above, the solution provided by the embodiments of the present invention can be applied to the case where there are at least two cover images of the work to be recommended. For each cover image, based on the content feature information of the work to be recommended, the cover image feature information of the cover image, and the target user feature information, the target feature information corresponding to the cover image is generated. Then, the target feature information is processed to obtain an evaluation score and determine the recommended work.

[0172] See Figure 5 The present invention provides a schematic diagram of the structure of an electronic device, including a processor 501, a communication interface 502, a memory 503 and a communication bus 504, wherein the processor 501, the communication interface 502 and the memory 503 communicate with each other through the communication bus 504.

[0173] Memory 503 is used to store computer programs;

[0174] When the processor 501 executes the program stored in the memory 503, it implements any of the steps in the aforementioned recommended work determination model training method.

[0175] As can be seen from the above, in the solution provided by the embodiments of the present invention, the parameters of the recommended work determination model are continuously adjusted through each training iteration, ultimately resulting in a well-trained recommended work determination model. In the recommended work determination model, by processing sample feature information, the content score and cover image score of the sample work are obtained in one step. Then, based on the content score and cover image score, an evaluation score is obtained for the sample work. Subsequently, sample works whose evaluation scores meet preset conditions can be determined as recommended works. The sample feature information includes the content feature information and cover image feature information of the sample work. In the process of calculating the evaluation score of the sample work, both the content and cover image of the sample work are considered simultaneously. Compared with considering only the content or cover image of the sample work, this provides a more comprehensive evaluation of the value of the sample work, improving the recommendation quality of the subsequently determined recommended works.

[0176] Furthermore, by comprehensively considering the characteristics of sample users, the content of sample works, and cover images, the recommended works determined based on evaluation scores are more applicable to real-world scenarios and can more accurately and realistically identify the works preferred by sample users. Analyzing the preferences of sample users can provide auxiliary decision-making analysis for the content platform, offering a basis for improving related products and services.

[0177] See Figure 6 This is a schematic diagram of another electronic device provided in an embodiment of the present invention, including a processor 601, a communication interface 602, a memory 603 and a communication bus 604, wherein the processor 601, the communication interface 602 and the memory 603 communicate with each other through the communication bus 604.

[0178] Memory 603 is used to store computer programs;

[0179] When the processor 601 executes the program stored in the memory 603, it implements any of the steps in the aforementioned method for determining recommended works.

[0180] As can be seen from the above, in the solution provided by the embodiments of the present invention, the above-mentioned target feature information is input into a pre-trained recommendation work determination model to obtain the evaluation score of the work to be recommended, and then the evaluation score can be used to determine whether the work to be recommended is recommended to the target user.

[0181] Furthermore, since the aforementioned target feature information includes the content and cover image features of the works to be recommended, the evaluation score obtained using the aforementioned work recommendation determination model considers both the content and cover image of the works to be recommended. Compared to considering only the content or cover image, this approach provides a more comprehensive assessment of the value of the works to be recommended, thus improving the quality of the final recommended works. By comprehensively considering the features of the target user, the content of the works to be recommended, and the cover image, the final recommended works are more applicable to real-world scenarios, more accurately and realistically identifying the works that target users like. This enhances the viewing experience for target users, increases the time they spend on the platform and their retention rate, and ultimately improves the click-through rate of advertisements on the platform. Different recommended works are determined for different target users, enriching personalized recommendation strategies and allowing advertisers to segment target customer groups. By comprehensively considering the content and cover image of the works to be recommended, resources can be used more rationally, resource allocation priorities can be adjusted, and resource utilization efficiency can be improved.

[0182] In addition, by using the above-mentioned recommended works determination model to consider both the content and cover image of the works, the resources consumed in calculating the evaluation scores of the works and thus determining the recommended works can be saved; the model can be applied to different fields that consider both the content and cover image of the works, and has scalability.

[0183] The communication bus mentioned in the above electronic devices can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This communication bus can be divided into address bus, data bus, control bus, etc. For ease of illustration, only one thick line is used to represent it in the diagram, but this does not indicate that there is only one bus or one type of bus.

[0184] The communication interface is used for communication between the aforementioned terminal and other devices.

[0185] The memory may include random access memory (RAM) or non-volatile memory, such as at least one disk storage device. Optionally, the memory may also be at least one storage device located remotely from the aforementioned processor.

[0186] The processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0187] In another embodiment of the present invention, a computer-readable storage medium is also provided, which stores a computer program that, when executed by a processor, implements the steps of the above-described recommended works determination model training method.

[0188] When training the recommended works determination model using the computer program stored in the computer-readable storage medium provided in this embodiment of the invention, the parameters of the recommended works determination model are continuously adjusted through each training session, ultimately resulting in a well-trained recommended works determination model. In the recommended works determination model, by processing sample feature information, the content score and cover image score of the sample works are obtained in one step. Then, based on the content score and cover image score, an evaluation score is obtained for the sample works. Sample works whose evaluation scores meet preset conditions can then be determined as recommended works. The sample feature information includes the content feature information and cover image feature information of the sample works. In the process of calculating the evaluation score of the sample works, both the content and cover image of the sample works are considered simultaneously. Compared with considering only the content or cover image of the sample works, this provides a more comprehensive evaluation of the value of the sample works and improves the recommendation quality of the subsequently determined recommended works.

[0189] Furthermore, by comprehensively considering the characteristics of sample users, the content of sample works, and cover images, the recommended works determined based on evaluation scores are more applicable to real-world scenarios and can more accurately and realistically identify the works preferred by sample users. Analyzing the preferences of sample users can provide auxiliary decision-making analysis for the content platform, offering a basis for improving related products and services.

[0190] In another embodiment of the present invention, a different computer-readable storage medium is also provided, which stores a computer program that, when executed by a processor, implements the steps of any of the above-described methods for determining recommended works.

[0191] When determining recommended works using the computer program stored in the computer-readable storage medium provided in this embodiment of the invention, the aforementioned target feature information is input into a pre-trained recommended work determination model to obtain an evaluation score for the work to be recommended. Then, based on the evaluation score, it can be determined whether the work to be recommended is recommended to the target user.

[0192] Furthermore, since the aforementioned target feature information includes the content and cover image features of the works to be recommended, the evaluation score obtained using the aforementioned work recommendation determination model considers both the content and cover image of the works to be recommended. Compared to considering only the content or cover image, this approach provides a more comprehensive assessment of the value of the works to be recommended, thus improving the quality of the final recommended works. By comprehensively considering the features of the target user, the content of the works to be recommended, and the cover image, the final recommended works are more applicable to real-world scenarios, more accurately and realistically identifying the works that target users like. This enhances the viewing experience for target users, increases the time they spend on the platform and their retention rate, and ultimately improves the click-through rate of advertisements on the platform. Different recommended works are determined for different target users, enriching personalized recommendation strategies and allowing advertisers to segment target customer groups. By comprehensively considering the content and cover image of the works to be recommended, resources can be used more rationally, resource allocation priorities can be adjusted, and resource utilization efficiency can be improved.

[0193] In addition, by using the above-mentioned recommended works determination model to consider both the content and cover image of the works, the resources consumed in calculating the evaluation scores of the works and thus determining the recommended works can be saved; the model can be applied to different fields that consider both the content and cover image of the works, and has scalability.

[0194] In another embodiment of the present invention, a computer program product containing instructions is also provided, which, when run on a computer, causes the computer to execute any of the recommended works determination model training methods described in the above embodiments.

[0195] When training the recommended works determination model using the computer program product provided in this embodiment of the invention, the parameters of the recommended works determination model are continuously adjusted through each training session, ultimately resulting in a well-trained recommended works determination model. In the recommended works determination model, by processing sample feature information, the content score and cover image score of the sample works are obtained in one step. Then, based on the content score and cover image score, an evaluation score is obtained for the sample works. Sample works whose evaluation scores meet preset conditions can then be determined as recommended works. The sample feature information includes the content feature information and cover image feature information of the sample works. In the process of calculating the evaluation score of sample works, both the content and cover image of the sample works are considered simultaneously. Compared with considering only the content or cover image of the sample works, this provides a more comprehensive evaluation of the value of the sample works, improving the recommendation quality of the subsequently determined recommended works.

[0196] Furthermore, by comprehensively considering the characteristics of sample users, the content of sample works, and cover images, the recommended works determined based on evaluation scores are more applicable to real-world scenarios and can more accurately and realistically identify the works preferred by sample users. Analyzing the preferences of sample users can provide auxiliary decision-making analysis for the content platform, offering a basis for improving related products and services.

[0197] In yet another embodiment of the present invention, another computer program product containing instructions is also provided, which, when run on a computer, causes the computer to perform any of the recommended works determination methods described above.

[0198] When using the computer program product provided in this embodiment of the invention to determine recommended works, the above-mentioned target feature information is input into a pre-trained recommended work determination model to obtain the evaluation score of the work to be recommended. Then, based on the evaluation score, it can be determined whether the work to be recommended is recommended to the target user.

[0199] Furthermore, since the aforementioned target feature information includes the content and cover image features of the works to be recommended, the evaluation score obtained using the aforementioned work recommendation determination model considers both the content and cover image of the works to be recommended. Compared to considering only the content or cover image, this approach provides a more comprehensive assessment of the value of the works to be recommended, thus improving the quality of the final recommended works. By comprehensively considering the features of the target user, the content of the works to be recommended, and the cover image, the final recommended works are more applicable to real-world scenarios, more accurately and realistically identifying the works that target users like. This enhances the viewing experience for target users, increases the time they spend on the platform and their retention rate, and ultimately improves the click-through rate of advertisements on the platform. Different recommended works are determined for different target users, enriching personalized recommendation strategies and allowing advertisers to segment target customer groups. By comprehensively considering the content and cover image of the works to be recommended, resources can be used more rationally, resource allocation priorities can be adjusted, and resource utilization efficiency can be improved.

[0200] In addition, by using the above-mentioned recommended works determination model to consider both the content and cover image of the works, the resources consumed in calculating the evaluation scores of the works and thus determining the recommended works can be saved; the model can be applied to different fields that consider both the content and cover image of the works, and has scalability.

[0201] In the above embodiments, implementation can be achieved entirely or partially through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented entirely or partially in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of the present invention are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid state disk (SSD)).

[0202] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, 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. Without further limitations, 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 said element.

[0203] The various embodiments in this specification are described in a related manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the embodiments of apparatus, electronic devices, computer-readable storage media, and computer program products are basically similar to the method embodiments, and therefore the descriptions are relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0204] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention are included within the scope of protection of the present invention.

Claims

1. A method for training a recommended works determination model, characterized in that, The method includes: Obtain historical behavior data of sample users; Extract the sample content feature information and sample cover image feature information of at least one sample work from the historical behavior data, and obtain the sample user feature information of the sample user; when there are multiple cover images of the sample work, obtain the sample cover image of the sample work in the following way: select a target number of sample cover images based on the weight factor of each cover image, wherein the cover image with the larger weight factor has a larger historical click rate and a higher probability of being selected; Based on the sample content feature information, sample cover image feature information, and sample user feature information, sample feature information is generated; The sample feature information is input into the recommended work determination model. The sample content score of the sample work is calculated based on the sample content feature information contained in the sample feature information. The sample cover image score is calculated based on the sample cover image feature information contained in the sample feature information. The sample content score and the sample cover image score are weighted according to their respective weights to calculate the evaluation score of the sample work. Based on the evaluation scores and the historical behavior data of the sample users, the sample loss of the recommended work determination model is calculated; Based on the sample loss, the parameters of the recommended work determination model are adjusted, and the step of inputting the sample feature information into the recommended work determination model is returned to be executed until the preset training convergence condition is reached, and the trained recommended work determination model is obtained.

2. The method according to claim 1, characterized in that, After selecting the target number of sample cover images, the process also includes: If the selected target number of sample cover images does not include the cover image shown to the sample user, the selected sample cover image is replaced with the shown cover image.

3. The method according to claim 1, characterized in that, After selecting the target number of sample cover images, the process also includes: If the selected target number of sample cover images does not include the cover image shown to the sample user, the shown cover image will be added to the sample cover images.

4. The method according to claim 1, characterized in that, The recommended work determination model contains multiple sub-networks, and the feature information of the sample cover images is processed in the following manner: The feature information of the sample cover image is processed by multiple sub-networks, and the processing results output by each sub-network are obtained. The processing results are combined and calculated based on the weights of each sub-network.

5. A method for determining recommended works, characterized in that, The method includes: Obtain the content feature information of the work to be recommended and the cover image feature information of the cover image of the work to be recommended; Obtain target user characteristic information; Based on the content feature information, cover image feature information, and target user feature information, target feature information is generated; The target feature information is input into a pre-trained recommended work determination model, which is trained according to the method steps of any one of claims 1-4; The output of the recommended work determination model is obtained to determine whether the work to be recommended is a recommended work.

6. The method according to claim 5, characterized in that, When the work to be recommended has at least two cover images, the step of generating target feature information based on the content feature information, cover image feature information, and target user feature information includes: For each of the at least two cover images, target feature information corresponding to the cover image is generated based on the content feature information, the cover image feature information of the cover image, and the target user feature information.

7. A recommended work determination model training device, characterized in that, The device includes: The first acquisition module is used to acquire historical behavior data of sample users; The extraction module is used to extract the sample content feature information of at least one sample work and the sample cover image feature information of the sample work from the historical behavior data, and to obtain the sample user feature information of the sample user; when there are multiple cover images of the sample work, the sample cover image of the sample work is obtained in the following way: based on the weight factor of each cover image, a target number of sample cover images are selected, wherein the cover image with the larger the weight factor has a larger historical click rate and a higher probability of being selected; The first generation module is used to generate sample feature information based on the sample content feature information, sample cover image feature information and sample user feature information; The input and output module is used to input the sample feature information into the recommended work determination model, calculate the sample content score of the sample work based on the sample content feature information contained in the sample feature information, calculate the sample cover image score based on the sample cover image feature information contained in the sample feature information, and calculate the evaluation score of the sample work by weighting the sample content score and the sample cover image score according to the set weights of their respective weights. The loss calculation module is used to calculate the sample loss of the recommended work determination model based on the evaluation score and the historical behavior data of the sample users. The adjustment module is used to adjust the parameters of the recommended work determination model based on the sample loss, and then return to the step of inputting the sample feature information into the recommended work determination model until the preset training convergence condition is reached, so as to obtain the trained recommended work determination model.

8. A device for determining recommended works, characterized in that, The device includes: The second acquisition module is used to acquire the content feature information of the work to be recommended and the cover image feature information of the cover image of the work to be recommended; The third acquisition module is used to acquire target user characteristic information of the target user; The second generation module is used to generate target feature information based on the content feature information, cover image feature information and target user feature information; The input module is used to input the target feature information into a pre-trained recommended works determination model, wherein the recommended works determination model is trained according to the method steps of any one of claims 1-4; The determination module is used to obtain the output result of the recommended work determination model and determine whether the work to be recommended is a recommended work.

9. An electronic device, characterized in that, It includes a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; Memory, used to store computer programs; A processor, when executing a program stored in memory, implements the method steps of any one of claims 1-4 or 5-6.

Images