Video recommendation model training method and device and video recommendation method and device

Abstract

The embodiment of the invention relates to a video recommendation model training method and device and a video recommendation method and device. The method comprises the steps of obtaining a video feature sample of an entry video of a first-level page and a current historical request feature sample of a second-level page; combining the video feature sample and the current historical request feature sample to obtain a cross feature sample; determining a next to-be-played video from a recalled video sample list based on a pre-constructed to-be-trained video recommendation model according to the cross feature sample, a plurality of video feature samples corresponding to the recalled video sample list and the user feature sample, and determining a loss value based on the similarity between the video feature sample of the next to-be-played video and the video feature sample of the next historically played video of the second-level page so as to adjust the model parameters to obtain a video recommendation model. According to the embodiment of the invention, the video recommendation model is trained through the cross feature of the entry video and the second-level page request information, so that the accuracy of video recommendation and the user experience are improved.

Description

Technical Field

[0001] The present disclosure relates to the field of multimedia processing technology, and in particular to a video recommendation model training, video recommendation method, device, equipment and medium. Background Art

[0002] At present, in the video recommendation scenario, there are a large number of users who jump to the video immersive page (referred to as the secondary page) through the video information flow page (referred to as the primary page); when the user selects an entry video on the primary page and enters the secondary page, the entry video can reflect the user's initial interest; at the same time, the user's immersive consumption process on the secondary page will reflect the user's interest continuation and transfer.

[0003] In the related art, if videos are recommended based on the entry video on the secondary page, it may be difficult to maintain the user's interest; if the user directly switches to the video corresponding to the user's current interest point after entering the secondary page, it may also lead to the failure of user interest transfer; therefore, how to effectively balance the relevance of the user's content consumption on the secondary page and the entry video on the primary page, while meeting the user's personalized consumption needs on the secondary page, is a key issue that needs to be solved. Summary of the invention

[0004] In order to solve the above technical problems or at least partially solve the above technical problems, the present disclosure provides a video recommendation model training, video recommendation method, device, equipment and medium.

[0005] The disclosed embodiment provides a method for training a video recommendation model, the method comprising: obtaining a video feature sample of an entry video of a first-level page and a current historical request feature sample of a second-level page; wherein, the second-level page is entered based on the entry video of the first-level page, and the current historical request feature sample of the second-level page refers to request information for requesting to play a video on the second-level page at the current historical time; the video feature sample and the current historical request feature sample are combined to obtain a cross-feature sample; multiple video feature samples corresponding to a recalled video sample list are obtained, and user feature samples are obtained; based on a pre-constructed video recommendation model to be trained, a next video to be played is determined from the recalled video sample list according to the cross-feature sample, the multiple video feature samples and the user feature sample, and a loss value is determined based on the similarity between the video feature sample of the next video to be played and the video feature sample of the next historical played video of the second-level page and a pre-set loss function to adjust the model parameters to obtain a video recommendation model.

[0006] Optionally, obtaining the current historical request feature sample of the secondary page includes: obtaining the current number of requests for the secondary page; obtaining the current viewed video parameters corresponding to the currently viewed video of the secondary page; wherein the currently viewed video parameters include at least one of a currently viewed video identifier, a currently viewed video tag, and a historically viewed video category; encoding the current number of requests, or encoding the current number of requests and the historically viewed video parameters to obtain the current historical request feature sample.

[0007] Optionally, the video feature sample and the current historical request feature sample are combined to obtain a cross-feature sample, including: multiplying the video feature sample and the current historical request feature sample to obtain the cross-feature sample; and / or concatenating the video feature sample and the current historical request feature sample to obtain the cross-feature sample; and / or obtaining weight information corresponding to the video feature sample and the current historical request feature sample, and processing the video feature sample and the current historical request feature sample based on the weight information to obtain the cross-feature sample.

[0008] Optionally, the method also includes: obtaining current viewing parameters of the currently viewed video on the secondary page; wherein the current viewing parameters include at least one of the current viewing time, the current viewing completion rate and the current switching frequency; and determining the weight information of the current historical request feature sample based on the current viewing parameters.

[0009] The disclosed embodiments also provide a video recommendation method, the method comprising: responding to an operation instruction for a target video on a primary page to play the target video on a secondary page; responding to a switch instruction for the secondary page to obtain video features of the target video, current request features of the secondary page, and user features; combining the video features and the current request features to obtain cross-features; based on a video recommendation model, searching the cross-features and the user features in a preset recall video list to obtain a recommended video to play on the secondary page; wherein the video recommendation model is obtained according to the training method of the video recommendation model described in the aforementioned embodiment.

[0010] Optionally, the video recommendation model retrieves the cross-features and the user features in a preset recalled video list to obtain recommended videos, including: obtaining candidate video features corresponding to each candidate video in the recalled video list; calculating based on the user features, the cross-features and the candidate video features through the video recommendation model to obtain target similarity; sorting according to all the target similarities, and obtaining the candidate video corresponding to the maximum target similarity as the recommended video.

[0011] The disclosed embodiment also provides a training device for a video recommendation model, comprising a first acquisition module for acquiring a video feature sample of an entry video of a first-level page; a second acquisition module for acquiring a current historical request feature sample of a second-level page; wherein the second-level page is entered based on the entry video of the first-level page, and the current historical request feature sample of the second-level page refers to request information for requesting to play a video on the second-level page at the current historical time; a first combination module for combining the video feature sample and the current historical request feature sample to obtain a cross-feature sample; a third acquisition module for acquiring a plurality of video feature samples and a user feature sample corresponding to a recalled video sample list; and a training module for determining a next video to be played from the recalled video sample list according to the cross-feature sample, the plurality of video feature samples and the user feature sample based on a pre-constructed video recommendation model to be trained, and determining a loss value based on the similarity between the video feature sample of the next video to be played and the video feature sample of the next historical played video of the second-level page and a pre-set loss function to adjust the model parameters to obtain a video recommendation model.

[0012] The disclosed embodiment also provides a video recommendation device, the method comprising: a response playback module, used to respond to an operation instruction for a target video on a primary page to play the target video on a secondary page; a response acquisition module, used to respond to a switch instruction for the secondary page to acquire video features of the target video, current request features of the secondary page and user features; a second combination module, used to combine the video features and the current request features to obtain cross features; a processing module, used to retrieve the cross features and the user features in a preset recall video list based on a video recommendation model to obtain a recommended video to play on the secondary page; wherein the video recommendation model is obtained according to the training method of the video recommendation model described in the aforementioned embodiment.

[0013] An embodiment of the present disclosure also provides an electronic device, comprising: a processor; a memory for storing executable instructions of the processor; the processor is used to read the executable instructions from the memory and execute the instructions to implement the training of the video recommendation model and the video recommendation method provided by the embodiment of the present disclosure.

[0014] The embodiments of the present disclosure also provide a computer-readable storage medium, which stores a computer program, and the computer program is used to execute the training of the video recommendation model and the video recommendation method provided by the embodiments of the present disclosure.

[0015] The embodiments of the present disclosure also provide a computer program product, including a computer program, wherein the computer program is executed by a processor to train a video recommendation model and a video recommendation method as provided in the embodiments of the present application.

[0016] The above-mentioned technical solution provided by the embodiment of the present disclosure combines the video features of the entry video and the request features of the secondary page for training the video recommendation model, so that the video recommendation model can mine the correlation behavior features between the primary page and the secondary page, so that based on the video recommendation model, it can realize the intelligent transition from the correlation between the secondary page recommended video and the entry video to the recommended video and the user's personalized interests, thereby improving the accuracy of video recommendations and user experience.

[0017] It should be understood that the content described in this section is not intended to identify the key or important features of the embodiments of the present disclosure, nor is it intended to limit the scope of the present disclosure. Other features of the present disclosure will become easily understood through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present disclosure and, together with the description, serve to explain the principles of the present disclosure.

[0019] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, for ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.

[0020] Figure 1 A schematic diagram of the training process of a video recommendation model provided for related technologies;

[0021] Figure 2 A flowchart of a method for training a video recommendation model provided in an embodiment of the present disclosure;

[0022] Figure 3 A schematic diagram of a primary page provided in an embodiment of the present disclosure;

[0023] Figure 4 A schematic diagram of a secondary page provided in an embodiment of the present disclosure;

[0024] Figure 5 A schematic diagram of a video recommendation process provided by an embodiment of the present disclosure;

[0025] Figure 6 A schematic diagram of the structure of a training device for a video recommendation model provided in an embodiment of the present disclosure;

[0026] Figure 7 A schematic diagram of the structure of a video recommendation device provided by an embodiment of the present disclosure;

[0027] Figure 8 A schematic diagram of the structure of an electronic device provided in an embodiment of the present disclosure. DETAILED DESCRIPTION

[0028] In order to more clearly understand the above-mentioned objectives, features and advantages of the present disclosure, the scheme of the present disclosure will be further described below. It should be noted that the embodiments of the present disclosure and the features in the embodiments can be combined with each other without conflict.

[0029] In the following description, many specific details are set forth to facilitate a full understanding of the present disclosure, but the present disclosure may also be implemented in other ways different from those described herein; it is obvious that the embodiments in the specification are only part of the embodiments of the present disclosure, rather than all of the embodiments.

[0030] In the existing video recommendation application scenario, a user selects an entry video through a video information flow page (referred to as the first-level page) and then jumps to a video immersive page (referred to as the second-level page). If videos related to the entry video are always recommended on the second-level page, it may be difficult to maintain the user's interest. Therefore, it is necessary to recommend videos to transfer the user's interest. If the user directly switches to the video corresponding to the user's current interest point after entering the second-level page, it may also lead to the failure of user interest transfer. Alternatively, the number of user requests is directly divided into relevance and personalization. For example, the first N requests on the second-level page are used to recommend videos based on the entry video, and subsequent personalized video recommendations are made based entirely on user interests. Among them, N is a positive integer. However, the selection of the request number threshold N is relatively inflexible, resulting in inaccurate transfer of video recommendation interest points, affecting the user experience.

[0031] In response to the above problems, when a user enters a secondary page from a primary page and consumes a video, in order to allow the user to transition more naturally from related videos to personalized interest videos, the disclosed embodiment proposes a training method for a video recommendation model, in which the user's request features on the secondary page are crossed with the video features of the entry video and used to train the video recommendation model. This achieves adaptive balanced optimization of the relevance and personalization of recommended videos based on the video recommendation model after the user enters the secondary page, thereby improving the user experience and meeting the user's needs.

[0032] Figure 1 The present invention provides a flow chart of a video recommendation model training method provided in an embodiment of the present invention. The method can be applied to an electronic device including a central processing unit, such as a computer, a mobile phone, a tablet computer, a television, a server, etc., without limitation. Figure 1As shown, the method mainly includes the following steps S102 to S108:

[0033] Step S102, obtain the video feature sample of the entry video of the first-level page and the current historical request feature sample of the second-level page; wherein, based on the entry video of the first-level page, the second-level page is entered, and the current historical request feature sample of the second-level page refers to the request information for requesting to play the video on the second-level page at the current historical time.

[0034] Among them, the entry video can be a video of any content. In the embodiment of the present disclosure, the video can be a short video; the first-level page refers to an information flow page including one or more entry videos, and the entry video is usually presented on the first-level page in the form of a static image or a dynamic image; the first-level page is usually the first page that a user sees after entering a video website or a video application, and the second-level page is the page entered after the first-level page. In the embodiment of the present disclosure, the second-level page is entered through the entry video in the first-level page, and the specific video content of the entry video is played on the second-level page. During the video playback on the second-level page, the next video can be requested to be played according to demand, such as switching to the next video playback page by sliding up. The current historical request feature sample of the second-level page refers to the request information for requesting to play a video on the second-level page at the current historical time, including the current number of requests, the video identifier corresponding to the currently viewed video, the video tag, the video category and other currently viewed video parameters.

[0035] It is understandable that, in the process of playing a video on a secondary page, multiple requests can be made to play the next video once, so the current historical time refers to the time of each request to play the next video.

[0036] In the embodiments of the present disclosure, there are many ways to obtain video feature samples of the entry video. As one example, the video identification of the entry video is obtained and encoded to obtain the video feature sample; as another example, the video identification, video category, and video tag of the entry video are obtained and encoded to obtain the video feature sample; wherein, according to the actual application needs, one or more of the video identification, video category, and video tag of the entry video can be selected for encoding to obtain the video feature sample of the entry video; wherein, the encoding can be performed through a neural network or an encoding model.

[0037] In the embodiments of the present disclosure, there are many ways to obtain the current historical request feature sample of the secondary page. As an example, the current number of requests for the secondary page is obtained and encoded to obtain the request feature sample; as another example, the current number of requests for the secondary page and the current viewed video parameters corresponding to the currently viewed video of the secondary page are obtained and encoded to obtain the request feature sample; wherein, the current number of requests for the secondary page and the current viewed video parameters corresponding to the currently viewed video of the secondary page can be selected according to actual application needs; wherein the current viewed video parameters include at least one of the currently viewed video identifier, the currently viewed video tag, and the currently viewed video category. For example, the current number of requests for the secondary page and one or more of the currently viewed video identifier, the currently viewed video tag, and the currently viewed video category are combined for encoding to obtain the current historical request feature sample of the secondary page.

[0038] Specifically, in response to the operation on the entry video of the first-level page, the second-level page is entered, the video feature sample of the entry video of the first-level page is obtained, the entry video is played on the second-level page, and the current number of requests for requesting to play the video in the current historical time, the video ID of the currently watched video, the video tag, etc. are obtained and encoded as the current historical request feature sample of the second-level page.

[0039] Step S104: Combine the video feature sample and the current historical request feature sample to obtain a cross feature sample.

[0040] Among them, the cross-feature sample refers to the feature sample of the combination of the video feature sample and the current historical request feature sample. Specifically, there are many video feature samples and there may also be many current historical request feature samples. Therefore, there may be multiple results of combining the video feature sample and the current historical request feature sample, that is, there may be multiple cross-feature samples.

[0041] In the embodiments of the present disclosure, there are many ways to combine the video feature sample and the current historical request feature sample to obtain a cross feature sample. As an example, the video feature sample and the current historical request feature sample are multiplied to obtain a cross feature sample; as another example, the video feature sample and the current historical request feature sample are concatenated to obtain a cross feature sample; as another example, the weight information corresponding to the video feature sample and the current historical request feature sample is obtained, and the video feature sample and the current historical request feature sample are processed based on the weight information to obtain a cross feature sample.

[0042] It should be noted that when there are multiple video feature samples or current historical request feature samples, some of the video feature samples and some of the current historical request feature samples can be multiplied, some of the video feature samples and some of the current historical request feature samples can be spliced, or some of the video feature samples and some of the current historical request feature samples can be calculated and processed according to weights, and all the obtained cross-feature samples can be used as cross-feature samples.

[0043] Specifically, after obtaining the video feature sample of the entry video of the first-level page and the current historical request feature sample of the second-level page, the video feature sample and the current historical request feature sample are combined to obtain a cross feature sample; wherein, there may be multiple cross feature samples.

[0044] Step S106, obtaining a plurality of video feature samples corresponding to the recalled video sample list, and obtaining a user feature sample.

[0045] Among them, the multiple video feature samples corresponding to the recalled video sample list refer to the video features corresponding to each video sample in the recalled video sample list, such as obtaining the visual features, subtitle text features, etc. corresponding to the video sample as the video feature samples of the video sample.

[0046] Among them, user feature samples refer to user features, including static features such as user's age, gender, etc., as well as features such as the user's recent video viewing history information and click video history information. User features can be determined through terminal identification and / or combined with user identification.

[0047] It should be noted that the execution order of steps 102 to 104 and step 106 can be determined according to actual conditions. Figure 1 These are examples only.

[0048] Step 108, based on the pre-built video recommendation model to be trained, the next video to be played is determined from the recalled video sample list according to the cross-feature samples, multiple video feature samples and user feature samples, and the loss value is determined based on the similarity between the video feature samples of the next video to be played and the video feature samples of the next historically played video on the secondary page and a pre-set loss function to adjust the model parameters to obtain the video recommendation model.

[0049] Among them, the next historical playback video of the secondary page is the video actually played in the history of the secondary page after the secondary page is requested to play at the current historical time; it should be noted that in the process of playing videos on the secondary page, the viewing time of each video will be recorded, and the historical playback videos with a viewing time greater than or equal to the preset time threshold will be used as positive samples for training, and the historical playback videos with a viewing time less than the preset time threshold will be used as negative samples for training the video recommendation model to be trained.

[0050] Among them, video recommendation includes video recall and video sorting. A relevant video recall model or algorithm can be used to pre-recall multiple video samples according to the user's historical viewing behavior and the video playback information, geographic location, network status, time information, etc. of the user's current application program, to obtain a recalled video sample list, and then sort the various video samples in the recalled video sample list, and finally recommend the video based on the sorting result. The embodiment of the present disclosure continuously adjusts the model parameters of the video recommendation model to be trained through the video features between the next video that has actually been played in history and the next video to be played determined by sorting the recalled video sample list during the model training process, to obtain a video recommendation model.

[0051] In the disclosed embodiment, the similarity between the cross-feature samples, the user feature samples and each video feature sample can be obtained, for example, the similarity is obtained by performing dot product of the cross-feature samples, the user feature samples and each video feature sample, and the cosine value between the cross-feature samples, the user feature samples and each video feature sample is calculated as the similarity, and the video feature sample corresponding to the maximum similarity is used as the target video feature sample, and the video sample corresponding to the target video feature sample is used as the next video to be played, and the similarity between the video feature sample of the next video to be played and the video feature sample of the next historically played video on the secondary page is input into the loss function of the pre-set model training to determine the loss value of this training, and the model parameters of the video recommendation model to be trained are adjusted based on the loss value, so that the model parameters of the video recommendation model to be trained can be continuously adjusted based on multiple cross-feature samples, user feature samples and multiple video feature samples, and finally the video recommendation model is obtained.

[0052] Specifically, the cross-feature samples, multiple video feature samples, and user feature samples can be processed by dot multiplication and the like to obtain the similarity between the cross-feature samples, the user feature samples, and each video feature sample, and the video sample corresponding to the maximum similarity is used as the next video to be played, and the similarity between the video feature sample of the next video to be played and the video feature sample of the next historically played video on the secondary page is input into a pre-set loss function for calculation to obtain a loss value. It can be understood that when the loss value is less than a certain threshold, the training can be stopped. If the loss value is greater than or equal to the threshold, the model parameters of the video recommendation model to be trained need to be adjusted, and then the model is continued to be trained until the loss value is less than the threshold to obtain a video recommendation model. In this way, the video recommendation model can be accurately and quickly obtained, thereby improving the accuracy of subsequent video recommendations.

[0053] It should be noted that, for the efficiency and effect of subsequent applications, multiple cross-feature sample input models can be trained and evaluated through AUC (Area Under Curve), and one or more target cross-feature samples can be determined from the multiple cross-feature samples based on the evaluation results. In subsequent practical applications, the video feature samples and request feature samples in the target cross-feature samples can be combined to obtain cross-features for retrieval, thereby further improving the accuracy of recommended videos.

[0054] Specifically, the predicted probabilities of the model are arranged in descending order; the predicted probabilities are used as thresholds from high to low, and the corresponding false positive rate FPR (False Positive Rate, the proportion of negative samples incorrectly classified as positive samples) and true positive rate TPR (True Positive Rate, the proportion of positive samples correctly classified as positive samples) are calculated; the calculated TPR and FPR values ​​are plotted in a two-dimensional coordinate system to form a ROC (Receiver Operating Characteristic) curve, and the area under the ROC curve is calculated using the numerical integration method to obtain AUC. The closer the AUC value is to 1, the better the performance of the model; the closer it is to 0.5, the worse the performance of the model.

[0055] In the above embodiment, the cross-feature samples obtained by combining the video feature samples and the request feature samples are used to train the video recommendation model. Since the cross-feature samples enhance the correlation between the primary page and the secondary page, the video recommendation results based on the video recommendation model are more accurate and comprehensive.

[0056] The training scheme of the video recommendation model provided by the embodiment of the present disclosure is through obtaining the video feature sample of the entry video of the first-level page and the current historical request feature sample of the second-level page; wherein, the second-level page is entered based on the entry video of the first-level page, and the current historical request feature sample of the second-level page refers to the request information for requesting to play the video on the second-level page at the current historical time; the video feature sample and the current historical request feature sample are combined to obtain a cross-feature sample; multiple video feature samples corresponding to the recalled video sample list are obtained, and user feature samples are obtained; based on the pre-constructed video recommendation model to be trained, the next video to be played is determined from the recalled video sample list according to the cross-feature sample, the multiple video feature samples and the user feature sample, and the loss value is determined based on the similarity between the video feature sample of the next video to be played and the video feature sample of the next historically played video of the second-level page and a pre-set loss function to adjust the model parameters to obtain the video recommendation model. By combining the video features of the entry video and the request features of the secondary page for training the video recommendation model, the video recommendation model can mine the associated behavioral features between the primary page and the secondary page. Therefore, based on the video recommendation model, the correlation between the secondary page recommended video and the entry video can be intelligently transitioned to the recommended video and the user's personalized interests, thereby improving the accuracy of video recommendations and user experience.

[0057] In some embodiments, obtaining a current historical request feature sample of a secondary page includes: obtaining the current number of requests for the secondary page; obtaining a current viewing video parameter corresponding to a currently viewed video of the secondary page; wherein the current viewing video parameter includes at least one of a currently viewed video identifier, a currently viewed video tag, and a historically viewed video category; encoding the current number of requests, or encoding the current number of requests and the historically viewed video parameters to obtain a current historical request feature sample.

[0058] Among them, the current number of requests for the secondary page refers to the number of requested video plays when the secondary page actively switches or automatically plays the next video after playing the video content of the entry video; the currently watched video refers to the video played on the secondary page at the current historical time.

[0059] Among them, the currently watched video identifier corresponding to the currently watched video refers to an identifier that can uniquely identify a currently watched video; the currently watched video tag corresponding to the currently watched video refers to a keyword related to the currently watched video, and the video can be pre-identified and analyzed in multiple dimensions such as scene, character, voice, and text based on deep learning and other methods to determine the label of the video; the currently watched video category corresponding to the currently watched video refers to a classification based on the content and form of the video, such as food, comedy, knowledge, and beauty, etc., and the specific setting is selected according to the actual application scenario.

[0060] In the present embodiment, one or more of the current number of requests for the secondary page, the current video identifier corresponding to the current video being viewed on the secondary page, the current video tag, the current video category, etc. can be selected as needed and encoded through a neural network or encoding model to obtain a current historical request feature sample of the secondary page.

[0061] In the above scheme, one or more of the current number of requests for the secondary page, the current viewed video identifier corresponding to the current viewed video of the secondary page, the current viewed video tag, the current viewed video category, etc. can be obtained and encoded as the current historical request feature sample of the secondary page, so that the precise features of the secondary page can be determined for combining subsequent cross-feature samples, and the cross-feature samples can be used to train the video recommendation model, thereby further improving the accuracy of video recommendations and user experience.

[0062] In some embodiments, the video feature sample and the current historical request feature sample are combined to obtain a cross-feature sample, including: multiplying the video feature sample and the current historical request feature sample to obtain a cross-feature sample; and / or concatenating the video feature sample and the current historical request feature sample to obtain a cross-feature sample; and / or obtaining weight information corresponding to the video feature sample and the current historical request feature sample, and processing the video feature sample and the current historical request feature sample based on the weight information to obtain a cross-feature sample.

[0063] In the disclosed embodiment, one or more methods may be selected to combine the video feature sample and the current historical request feature sample to obtain a cross feature sample.

[0064] As an example, the video feature sample includes: a video identification feature sample; the current historical request feature sample of the secondary page includes: a current request count feature sample, a current viewed video tag feature sample corresponding to the currently viewed video, and a current viewed video category feature sample corresponding to the currently viewed video, and the video identification feature sample and the current request count feature sample are dotted to obtain a cross feature sample; the video identification feature sample and the current viewed video tag feature sample corresponding to the currently viewed video are concatenated to obtain a cross feature sample; based on the weight information corresponding to the current viewed video category feature sample corresponding to the currently viewed video, a first weight corresponding to the current viewed video category feature sample corresponding to the currently viewed video is determined, and based on the target value such as 1 and the first weight, a second weight corresponding to the video identification feature sample is determined, thereby performing weighted calculation based on the video identification feature sample, the second weight, the current viewed video category feature sample corresponding to the currently viewed video, and the first weight to obtain a cross feature sample.

[0065] In some embodiments, in order to further improve the accuracy of cross-feature sample determination, the method also includes: obtaining current viewing parameters of the currently viewed video on the secondary page; wherein the current viewing parameters include at least one of the current viewing time, the current viewing completion rate and the current switching frequency; and determining the weight information of the current historical request feature sample based on the current viewing parameters.

[0066] Among them, the current viewing time of the currently watched video refers to the length of time each video is watched on the secondary page in the current historical time; the viewing completion rate of the currently watched video refers to the ratio of the watched video length of each video to the total video length when watching the video on the secondary page in the current historical time; the current switching frequency of the secondary page refers to the number of video switching operations per unit time when watching videos on the secondary page in the current historical time.

[0067] In an embodiment of the present disclosure, weight information of a request feature sample can be determined based on one or more of the current viewing time, the current viewing completion rate, and the current switching frequency of the currently viewed video on the secondary page. For example, if the viewing completion rate of a currently viewed video is 1, the weight of the current historical request feature sample determined based on the corresponding current viewed video identifier of the currently viewed video, and / or the current viewed video tag, and / or the current viewed video category can be set to 0.9; for another example, if the viewing completion rate of a currently viewed video is 0.1, the weight of the current historical request feature sample determined based on the corresponding current viewed video identifier of the currently viewed video, and / or the current viewed video tag, and / or the current viewed video category can be set to 0.

[0068] In the above scheme, one or more methods can be selected to combine video feature samples and current historical request feature samples to obtain cross-feature samples, thereby improving the diversity of cross-feature samples, and then using the cross-feature samples to train the video recommendation model to further improve the accuracy of video recommendation and user experience.

[0069] Figure 2 The present invention provides a flow chart of a video recommendation model training method, which mainly includes the following steps S202 to S208:

[0070] Step S202, obtain the video feature sample of the entry video of the first-level page, obtain the current number of requests for the second-level page, and obtain the current viewing video parameters corresponding to the viewing video of the second-level page; encode the current number of requests, or encode the current number of requests and the historical viewing video parameters to obtain the current historical request feature sample.

[0071] For example, Figure 3The first-level page shown includes Video 1, Video 2, Video 3, Video 4, Video 5, Video 6, Video 7 and Video 8, which are displayed in the form of a video list. Video 1 to Video 8 can be presented on the first-level page in the form of static images or dynamic images. Users can operate any video from Video 1 to Video 8 to enter the second-level page according to actual application needs. In other words, Video 1 to Video 8 can all be used as entry videos.

[0072] For example, Figure 3 Click on video 1 in the video to enter the secondary page, such as Figure 4 The secondary page shown plays the video content of Video 1. During the playback of Video 1 on the secondary page, the user can slide up or click the switch control as needed to play the next video. The next video can also be played automatically after the playback of Video 1 on the secondary page ends. The specific settings are selected according to the actual application scenario.

[0073] Specifically, in order to improve the user's immersive video viewing on the secondary page, so that the next video to be played, that is, the recommended video, better meets the user's interest needs, the video feature sample of the entry video of the primary page, the current number of requests of the secondary page, and the current viewing video parameters corresponding to the currently viewed video are obtained for encoding processing to obtain the current historical request feature samples for combination for training the video recommendation model; wherein the currently viewed video parameters include at least one of the currently viewed video identifier, the currently viewed video tag, and the historically viewed video category.

[0074] The video feature sample of the entry video refers to extracting video features from the entry video selected by the user on the first-level page, such as video identification, video category, video tag, etc., and encoding them to obtain the video feature sample.

[0075] Among them, the current historical request feature sample of the secondary page refers to extracting the current number of requests made by the user on the secondary page (that is, indicating the number of times the user has requested or refreshed the content) for encoding to obtain the current historical request feature sample as a sign of user interest decay or exploration behavior in the secondary page. The current number of requests can also be selected and combined with the current viewing video parameters corresponding to the current video being watched by the user on the secondary page, such as the current viewing video identifier, the current viewing video tag, the current viewing video category, etc., for encoding to obtain the current historical request feature sample.

[0076] Step S204, multiplying the video feature sample and the current historical request feature sample to obtain a cross feature sample; and / or, concatenating the video feature sample and the current historical request feature sample to obtain a cross feature sample; and / or, obtaining weight information corresponding to the video feature sample and the current historical request feature sample, and processing the video feature sample and the current historical request feature sample based on the weight information to obtain a cross feature sample.

[0077] In some embodiments, current viewing parameters of a currently viewed video on a secondary page are obtained; wherein the current viewing parameters include at least one of a current viewing duration, a current viewing completion rate, and a current switching frequency, and weight information of a current historical request feature sample is determined based on the current viewing parameters.

[0078] Specifically, the video feature samples of the entry video of the first-level page are combined with the current historical request feature samples of the second-level page. When combining, the weight information of the current historical request feature samples is determined according to the current viewing parameters of the currently viewed video on the second-level page to generate a set of cross feature samples; for example, the video category feature samples of the entry video are crossed with the feature samples of the number of times the user has requested on the second-level page, and the current viewing time of the user in the content requested at that time is used as the weight to reflect the user's deep interest and continuous preference in this category.

[0079] It should be noted that different cross-feature sample combinations are designed with different cross-feature methods (such as concatenation, multiplication, attention mechanism, etc.) to reflect users’ continued interest in different types of content or their intention to explore new content, further improving the diversity of cross-feature samples and thus improving the accuracy of the video recommendation model.

[0080] It should be noted that the effectiveness of cross-feature samples can also be evaluated in a targeted manner, and cross-feature samples with significant recommendation effects can be retained to further ensure the effectiveness of cross-feature samples and improve the processing effect of the video recommendation model.

[0081] Step S206, obtaining a plurality of video feature samples corresponding to the recalled video sample list, and obtaining a user feature sample.

[0082] It should be noted that step 206 is the same as step 106. For details, please refer to the detailed description of step 106, which will not be described in detail here.

[0083] Step S208, based on the pre-built video recommendation model to be trained, the next video to be played is determined from the recalled video sample list according to the cross-feature samples, multiple video feature samples and user feature samples, and the loss value is determined based on the similarity between the video feature samples of the next video to be played and the video feature samples of the next historically played video on the secondary page and a pre-set loss function to adjust the model parameters to obtain a video recommendation model.

[0084] Specifically, cross-feature samples are introduced into the bottom layer of the deep neural network side of the video recommendation model to be trained, so that the video recommendation model to be trained can capture the correlation between the video feature samples of the entry video and the current historical request feature samples at the initial layer of the deep neural network, thereby better understanding the user's interest dynamics under different request information.

[0085] It should be noted that the user's historical data is used for training, and the effect of the video recommendation model is tested through the validation set to ensure the relevance and personalized adaptability of the recommended videos under different request information.

[0086] Specifically, the video feature samples of the entry video of the first-level page and the request feature samples of the second-level page are cross-introduced into the video recommendation model to be trained to capture the correlation between the entry video features and the request information, so as to better understand the continuation and transfer of user interests, and the video identification, video tag, video category and other features of the entry video can be crossed with the current number of requests of the second-level page and added to the video recommendation model to be trained for training, so that the generated video recommendation model can explore the user's interest dynamics under different request information. By introducing the cross-features of the entry video and the request information, combined with the deep learning of the video recommendation model to be trained, the balance adaptability of the recommended video between relevance and personalization is improved, and the consumption time and user experience of users on the second-level page of the video scene are further improved.

[0087] Therefore, by collecting video feature samples of the entry video of users entering the secondary page from the first-level page in historical time, as well as the request feature samples of users on the secondary page, a cross-address method is adopted to add them to the video recommendation model to be trained for training. The video recommendation model obtained by training can mine the correlation behavior characteristics between users on the first-level page and the second-level page, thereby realizing the intelligent transition from the correlation between the recommended video on the secondary page and the entry video to the recommended video and the user's personalized interests, thereby improving the accuracy of the recommendation and the user experience.

[0088] Figure 5 The following is a flow chart of a video recommendation method provided by an embodiment of the present disclosure. The method mainly includes the following steps S302 to S310:

[0089] Step S302, responding to an operation instruction on a target video on a primary page, to play the target video on a secondary page.

[0090] Step S304, in response to the switching instruction to the secondary page, obtaining the video features of the target video, the current request features of the secondary page, and the user features.

[0091] Step S306: Combine the video feature and the current request feature to obtain a cross feature.

[0092] Step S308, based on the video recommendation model, searching the preset recalled video list according to the cross features and the user features, and obtaining the recommended video to be played on the secondary page.

[0093] In the embodiment of the present disclosure, the video recommendation model is obtained by the training method of the video recommendation model described in the above embodiment.

[0094] In the actual short video playback scenario, the recalled video list can be determined in advance, and the videos in the recalled video list can be sorted so that they can be played according to the sorting result when the user requests to play the next video on the secondary page. In this embodiment, the video recommendation model obtained through training can mine the user's associated behavioral characteristics between the primary page and the secondary page, so that the videos in the recalled video list can be accurately sorted, thereby recommending the video to be played on the secondary page.

[0095] In this embodiment, the target video is a target video triggered by the user on the primary page according to the actual video viewing needs, and the target video is played on the secondary page, thereby obtaining the video features of the target video and the current request feature combination of the secondary page, obtaining the cross-features, and inputting the cross-features and user features into the video recommendation model to perform video retrieval in the recalled video list, and the recommended video can be obtained; the recommended video refers to the next video requested to be played on the secondary page.

[0096] As a possible implementation method, the candidate video features corresponding to each candidate video in the recalled video list are obtained, and the target similarity is calculated between the cross features and the user features and the candidate video features corresponding to each candidate video in the recalled video list, and the candidate video corresponding to the target similarity greater than the preset similarity threshold is used as the recommended video (if there are multiple, one can be randomly selected as the recommended video), or the candidate video corresponding to the maximum target similarity is used as the recommended video, etc. Among them, the candidate video can be a video clip to meet the video recommendation needs of users in different scenarios.

[0097] Specifically, when a user initiates a video switching request, the cross-features and user features obtained in real time are input into the video recommendation model to return the most relevant recommended video.

[0098] Therefore, in the process of watching videos on the secondary page, the recommended videos can be accurately obtained based on the video recommendation model, so that the recommended videos corresponding to the user's historical interest points can be switched to at the precise time point, realizing the intelligent transition from the correlation between the recommended videos on the secondary page and the entry video to the recommended videos and the user's personalized interests, thereby improving the accuracy of recommendations and user experience.

[0099] Corresponding to the aforementioned video recommendation model training method, the present disclosure further provides a video recommendation model training device. Figure 6 This is a schematic diagram of the structure of a video recommendation model training device provided in an embodiment of the present disclosure. The device can be implemented by software and / or hardware and is applied to an electronic device including a central processing unit. The device includes:

[0100] The first acquisition module 402 is used to acquire a video feature sample of an entry video of a first-level page;

[0101] The second acquisition module 404 is used to acquire the current historical request feature sample of the secondary page; wherein, the secondary page is entered based on the entry video of the primary page, and the current historical request feature sample of the secondary page refers to the request information for requesting to play a video on the secondary page at the current historical time;

[0102] A first combining module 406, configured to combine the video feature sample and the current historical request feature sample to obtain a cross feature sample;

[0103] The third acquisition module 408 is used to acquire multiple video feature samples and user feature samples corresponding to the recalled video sample list;

[0104] The training module 410 is used to determine the next video to be played from the recalled video sample list based on the pre-built video recommendation model to be trained according to the cross-feature samples, the multiple video feature samples and the user feature samples, and determine the loss value based on the similarity between the video feature sample of the next video to be played and the video feature sample of the next historically played video of the secondary page and a pre-set loss function to adjust the model parameters to obtain a video recommendation model.

[0105] The above-mentioned device provided by the embodiment of the present disclosure combines the video features of the entry video and the request features of the secondary page for training the video recommendation model, so that the video recommendation model can mine the associated behavioral features between the primary page and the secondary page, so that based on the video recommendation model, it can realize the intelligent transition from the correlation between the secondary page recommended video and the entry video to the recommended video and the user's personalized interests, thereby improving the accuracy of video recommendations and user experience.

[0106] In some embodiments, the second acquisition module 404 is specifically used to: obtain the current number of requests for the secondary page; obtain the current viewing video parameters corresponding to the currently viewed video of the secondary page; wherein the currently viewed video parameters include at least one of the currently viewed video identifier, the currently viewed video tag, and the historically viewed video category; encode the current number of requests, or encode the current number of requests and the historically viewed video parameters to obtain the current historical request feature sample.

[0107] In some embodiments, the first combination module is specifically used to: perform multiplication processing on the video feature sample and the current historical request feature sample to obtain the cross feature sample; and / or, perform splicing processing on the video feature sample and the current historical request feature sample to obtain the cross feature sample; and / or, obtain weight information corresponding to the video feature sample and the current historical request feature sample, and process the video feature sample and the current historical request feature sample based on the weight information to obtain the cross feature sample.

[0108] In some embodiments, the device also includes: a fourth acquisition module, used to obtain current viewing parameters of the currently viewed video on the secondary page; wherein the current viewing parameters include at least one of the current viewing time, the current viewing completion rate and the current switching frequency; a determination module, used to determine the weight information of the current historical request feature sample based on the current viewing parameters.

[0109] The video recommendation model training device provided in the embodiments of the present disclosure can execute the video recommendation model training method provided in any embodiment of the present disclosure, and has the corresponding functional modules and beneficial effects of the execution method.

[0110] Corresponding to the aforementioned video recommendation method, the embodiment of the present disclosure further provides a video recommendation device, Figure 7 This is a schematic diagram of the structure of a video recommendation device provided by an embodiment of the present disclosure. The device can be implemented by software and / or hardware, and the device includes:

[0111] A response playing module 502, for responding to an operation instruction on a target video on a primary page to play the target video on a secondary page;

[0112] A response acquisition module 504, configured to acquire, in response to a switch instruction to the secondary page, video features of the target video, current request features of the secondary page, and user features;

[0113] A second combining module 506, configured to combine the video feature and the current request feature to obtain a cross feature;

[0114] The processing module 508 is used to retrieve the cross-features and the user features in a preset recalled video list based on a video recommendation model to obtain a recommended video for playing on the secondary page; wherein the video recommendation model is obtained according to the training method of the video recommendation model described in the aforementioned embodiment.

[0115] In some embodiments, the processing module 508 is specifically used to obtain candidate video features corresponding to each candidate video in the recalled video list; calculate based on the user features, the cross features and the candidate video features through the video recommendation model to obtain the target similarity; sort according to all the target similarities, and obtain the candidate video corresponding to the maximum target similarity as the recommended video.

[0116] The above-mentioned device provided by the embodiment of the present disclosure can accurately obtain recommended videos based on the video recommendation model during the process of watching videos on the secondary page, so that it can switch to the recommended video corresponding to the user's historical interest point at a precise time point, and realize the intelligent transition from the correlation between the recommended video on the secondary page and the entry video to the recommended video and the user's personalized interests, thereby improving the accuracy of recommendations and user experience.

[0117] The video recommendation device provided in the embodiments of the present disclosure can execute the video recommendation method provided in any embodiment of the present disclosure, and has the corresponding functional modules and beneficial effects of the execution method.

[0118] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the above-described device embodiment can refer to the corresponding process in the method embodiment, and will not be repeated here.

[0119] An embodiment of the present disclosure provides an electronic device, which includes: a storage device on which a computer program is stored; and a processing device for executing the computer program in the storage device to implement the steps of any method in the present disclosure.

[0120] Reference below Figure 8 , which shows a schematic diagram of the structure of an electronic device 800 suitable for implementing the embodiment of the present disclosure. The terminal device in the embodiment of the present disclosure may include but is not limited to mobile terminals such as mobile phones, laptop computers, digital broadcast receivers, PDAs (personal digital assistants), PADs (tablet computers), PMPs (portable multimedia players), vehicle-mounted terminals (such as vehicle-mounted navigation terminals), etc., and fixed terminals such as digital TVs, desktop computers, etc. Figure 8 The electronic device shown is only an example and should not bring any limitation to the functions and scope of use of the embodiments of the present disclosure.

[0121] like Figure 8As shown, the electronic device 800 may include a processing device (e.g., a central processing unit, a graphics processing unit, etc.) 801, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 802 or a program loaded from a storage device 808 into a random access memory (RAM) 803. In the RAM 803, various programs and data required for the operation of the electronic device 800 are also stored. The processing device 801, the ROM 802, and the RAM 803 are connected to each other via a bus 804. An input / output (I / O) interface 805 is also connected to the bus 804.

[0122] Typically, the following devices may be connected to the I / O interface 805: input devices 806 including, for example, a touch screen, a touch pad, a keyboard, a mouse, a camera, a microphone, an accelerometer, a gyroscope, etc.; output devices 807 including, for example, a liquid crystal display (LCD), a speaker, a vibrator, etc.; storage devices 808 including, for example, a magnetic tape, a hard disk, etc.; and communication devices 809. The communication device 809 may allow the electronic device 800 to communicate with other devices wirelessly or by wire to exchange data. Although Figure 8 The electronic device 800 is shown with various devices, but it should be understood that it is not required to implement or possess all the devices shown. More or fewer devices may be implemented or possessed instead.

[0123] In particular, according to an embodiment of the present disclosure, the process described above with reference to the flowchart can be implemented as a computer software program. For example, an embodiment of the present disclosure includes a computer program product, which includes a computer program carried on a non-transitory computer-readable medium, and the computer program contains program code for executing the method shown in the flowchart. In such an embodiment, the computer program can be downloaded and installed from a network through a communication device 809, or installed from a storage device 808, or installed from a ROM 802. When the computer program is executed by the processing device 801, the above-mentioned functions defined in the method of the embodiment of the present disclosure are executed.

[0124] In addition to the above-mentioned methods and devices, the embodiments of the present disclosure may also be a computer program product, which includes computer program instructions, which, when executed by a processor, cause the processor to perform the image processing method provided by the embodiments of the present disclosure. The computer program product may be written in any combination of one or more programming languages ​​to write program codes for performing the operations of the embodiments of the present disclosure, the programming languages ​​including object-oriented programming languages ​​such as Java, C++, etc., and also conventional procedural programming languages ​​such as "C" language or similar programming languages. The program code may be executed entirely on a user computing device, partially on a user device, as an independent software package, partially on a user computing device and partially on a remote computing device, or entirely on a remote computing device or server.

[0125] In addition, the embodiment of the present disclosure may also be a computer-readable storage medium on which computer program instructions are stored. When the computer program instructions are executed by a processor, the processor executes the training method of the video recommendation model provided by the embodiment of the present disclosure.

[0126] The computer readable storage medium can adopt any combination of one or more readable media. The readable medium can be a readable signal medium or a readable storage medium. The readable storage medium can include, for example, but is not limited to, a system, device or device of electricity, magnetism, light, electromagnetic, infrared, or semiconductor, or any combination of the above. More specific examples (non-exhaustive list) of readable storage media include: an electrical connection with one or more wires, a portable disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above.

[0127] The embodiments of the present disclosure also provide a computer program product, including a computer program / instruction, which, when executed by a processor, implements the training method of the video recommendation model in the embodiments of the present disclosure.

[0128] It is understandable that before using the technical solutions disclosed in the various embodiments of the present disclosure, the types, scope of use, usage scenarios, etc. of the personal information involved in the present disclosure should be informed to the user and the user's authorization should be obtained in an appropriate manner in accordance with relevant laws and regulations.

[0129] For example, in response to receiving an active request from a user, a prompt message is sent to the user to clearly prompt the user that the operation requested to be performed will require obtaining and using the user's personal information. Thus, the user can autonomously choose whether to provide personal information to software or hardware such as an electronic device, application, server, or storage medium that performs the operation of the technical solution of the present disclosure according to the prompt message.

[0130] As an optional but non-limiting implementation, in response to receiving an active request from the user, the prompt information may be sent to the user in the form of a pop-up window, in which the prompt information may be presented in text form. In addition, the pop-up window may also carry a selection control for the user to choose "agree" or "disagree" to provide personal information to the electronic device.

[0131] It is understandable that the above notification and the process of obtaining user authorization are merely illustrative and do not constitute a limitation on the implementation of the present disclosure. Other methods that meet relevant laws and regulations may also be applied to the implementation of the present disclosure.

[0132] It should be noted that, in this article, relational terms such as "first" and "second" are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the existence of other identical elements in the process, method, article or device including the elements.

[0133] The above description is only a specific embodiment of the present disclosure, so that those skilled in the art can understand or implement the present disclosure. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present disclosure. Therefore, the present disclosure will not be limited to the embodiments described herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A video recommendation model training method, characterized in that: include: Obtaining a video feature sample of the entry video of the first-level page and a current historical request feature sample of the second-level page; wherein, the second-level page is entered based on the entry video of the first-level page, and the current historical request feature sample of the second-level page refers to the request information for requesting to play a video on the second-level page at the current historical time; Combining the video feature sample and the current historical request feature sample to obtain a cross feature sample; Obtain multiple video feature samples corresponding to the recalled video sample list, and obtain user feature samples; Based on the pre-built video recommendation model to be trained, the next video to be played is determined from the recalled video sample list according to the cross-feature samples, the multiple video feature samples and the user feature samples, and the loss value is determined based on the similarity between the video feature sample of the next video to be played and the video feature sample of the next historically played video of the secondary page and a pre-set loss function to adjust the model parameters to obtain a video recommendation model.

2. The method according to claim 1, characterized in that Obtaining the current historical request feature sample of the secondary page, including: Get the current number of requests for the secondary page; Acquire a currently viewed video parameter corresponding to the currently viewed video on the secondary page; wherein the currently viewed video parameter includes at least one of a currently viewed video identifier, a currently viewed video tag, and a historically viewed video category; The current number of requests is encoded, or the current number of requests and the historical video viewing parameters are encoded to obtain the current historical request feature sample.

3. The method according to claim 1, characterized in that: The combining the video feature sample and the current historical request feature sample to obtain a cross feature sample includes: Performing product processing on the video feature sample and the current historical request feature sample to obtain the cross feature sample; and / or, Performing splicing processing on the video feature sample and the current historical request feature sample to obtain the cross feature sample; and / or, Obtain weight information corresponding to the video feature sample and the current historical request feature sample, and process the video feature sample and the current historical request feature sample based on the weight information to obtain the cross feature sample.

4. The method according to claim 3, characterized in that The method further comprises: Acquire the current viewing parameters of the currently viewed video on the secondary page; wherein the current viewing parameters include at least one of the current viewing duration, the current viewing completion rate, and the current switching frequency; The weight information of the current historical request feature sample is determined based on the current viewing parameter.

5. A video recommendation method, characterized in that: include: In response to an operation instruction on a target video on the primary page, the target video is played on the secondary page; In response to a switch instruction to the secondary page, obtaining video features of the target video, current request features of the secondary page, and user features; Combining the video feature and the current request feature to obtain a cross feature; Based on the video recommendation model, the cross-features and the user features are retrieved in a preset recalled video list to obtain a recommended video for playing on the secondary page; wherein the video recommendation model is obtained according to the training method of the video recommendation model described in any one of claims 1-4.

6. The method according to claim 5, characterized in that The video recommendation model retrieves the cross feature and the user feature in a preset recalled video list to obtain a recommended video, including: Obtain candidate video features corresponding to each candidate video in the recalled video list; The video recommendation model is used to calculate based on the user features, the cross features and the candidate video features to obtain a target similarity; All the target similarities are sorted, and the candidate video corresponding to the maximum target similarity is obtained as the recommended video.

7. A video recommendation model training device, characterized in that: include: A first acquisition module is used to acquire a video feature sample of an entry video of a first-level page; The second acquisition module is used to acquire the current historical request feature sample of the secondary page; wherein, the secondary page is entered based on the entry video of the primary page, and the current historical request feature sample of the secondary page refers to the request information for requesting to play a video on the secondary page at the current historical time; A first combining module, used for combining the video feature sample and the current historical request feature sample to obtain a cross feature sample; A third acquisition module is used to acquire multiple video feature samples and user feature samples corresponding to the recalled video sample list; A training module is used to determine the next video to be played from the recalled video sample list based on the pre-built video recommendation model to be trained according to the cross-feature samples, the multiple video feature samples and the user feature samples, and determine the loss value based on the similarity between the video feature sample of the next video to be played and the video feature sample of the next historically played video of the secondary page and a pre-set loss function to adjust the model parameters to obtain a video recommendation model.

8. A video recommendation device, characterized in that: include: A response playing module, used for responding to an operation instruction on a target video on a primary page to play the target video on a secondary page; A response acquisition module, configured to acquire, in response to a switching instruction to the secondary page, video features of the target video, current request features of the secondary page, and user features; A second combining module, used for combining the video feature and the current request feature to obtain a cross feature; A processing module is used to retrieve the cross-features and the user features in a preset recalled video list based on a video recommendation model to obtain a recommended video for playing on the secondary page; wherein the video recommendation model is obtained according to the training method of the video recommendation model described in any one of claims 1-4.

9. An electronic device, characterized in that: The electronic device comprises: a storage device having a computer program stored thereon; A processing device, used to execute the computer program in the storage device to implement the steps of the video recommendation model training method described in any one of claims 1 to 4, or the steps of the video recommendation method described in claim 5 or 6.

10. A computer-readable storage medium, characterized in that: The storage medium stores a computer program, and the computer program is used to execute the steps of the video recommendation model training method described in any one of claims 1 to 4, or the steps of the video recommendation method described in claim 5 or 6.

Images