Video recommendation method and device, model training method and device, equipment and storage medium

Abstract

The invention provides a video recommendation method and device, a model training method and device, equipment and a storage medium, and relates to deep learning, data processing and recommendation algorithms in the technical field of artificial intelligence. According to the specific implementation scheme, historical watching data corresponding to a plurality of users are calibrated based on labels associated with user preferences, a plurality of training samples are obtained, iterative training is conducted on a multi-target model according to the training samples and the labels, and a fine-arrangement multi-target model used for predicting the recommendation index of a video to a specific user is obtained. And according to the historical watching data of the target user, predicting the candidate videos by adopting a fine arrangement multi-target model to obtain recommendation indexes of the candidate videos, sorting the candidate videos according to the recommendation indexes to obtain a recommendation list, and pushing the recommendation list to the target user. According to the method, more flexible, accurate and personalized video recommendation is realized, and the operation complexity of querying and retrieving the preferred video by the user is reduced.

Description

Technical Field

[0001] The present disclosure relates to deep learning, data processing, and recommendation algorithms in the field of artificial intelligence technology, and in particular to a video recommendation method, model training method, device, equipment, and storage medium. Background Art

[0002] With the rapid development of mobile Internet technology, the penetration rate of smart phones has increased significantly, providing a solid foundation for the rise of short video applications. Users can watch, create and share short videos through their mobile phones anytime and anywhere, which greatly improves the dissemination efficiency of content and user participation. In the fast-paced life, people are increasingly inclined to obtain entertainment and relaxation through short videos. Short videos meet users' fragmented entertainment needs with their concise and rich content.

[0003] Nowadays, most short video applications usually recommend videos to users by analyzing the content features of videos through filtering algorithms.

[0004] However, how to recommend videos to users more flexibly and accurately and improve user experience is a technical problem in this field. Summary of the invention

[0005] The present disclosure provides a video recommendation method, a model training method, an apparatus, a device, and a storage medium.

[0006] According to a first aspect of the present disclosure, a video recommendation method is provided, comprising:

[0007] For each candidate video in the pre-acquired candidate video set, a pre-trained precision ranking multi-objective model is used to predict the candidate video according to the historical viewing data of the target user to obtain a recommendation index of the candidate video, wherein the precision ranking multi-objective model is a model trained based on pre-set labels associated with user preferences and used to predict the recommendation index of a video for a specific user;

[0008] Sorting the candidate videos in the candidate video set from large to small according to the recommendation index to obtain a recommendation list;

[0009] The recommendation list is pushed to the target user.

[0010] According to a second aspect of the present disclosure, a model training method is provided, comprising:

[0011] Calibrate historical viewing data corresponding to a plurality of users based on tags associated with user preferences to obtain a plurality of training samples, wherein the plurality of training samples include positive samples indicating recommendation and negative samples indicating recommendation;

[0012] Dividing the multiple training samples into a training set and a test set;

[0013] The pre-built multi-objective model is iteratively trained according to the training set, the test set and the labels to obtain a refined multi-objective model for predicting the recommendation index of a video for a specific user.

[0014] According to a third aspect of the present disclosure, a video recommendation device is provided, comprising:

[0015] A prediction unit is used to predict each candidate video in the pre-acquired candidate video set according to the historical viewing data of the target user using a pre-trained precision ranking multi-objective model to obtain a recommendation index of the candidate video, wherein the precision ranking multi-objective model is a model trained based on pre-set labels associated with user preferences and used to predict the recommendation index of a video for a specific user;

[0016] A sorting unit, used to sort the candidate videos in the candidate video set from large to small according to the recommendation index to obtain a recommendation list;

[0017] A pushing unit is used to push the recommendation list to the target user.

[0018] According to a fourth aspect of the present disclosure, a model training device is provided, comprising:

[0019] A calibration unit, configured to calibrate historical viewing data corresponding to a plurality of users based on tags associated with user preferences to obtain a plurality of training samples, wherein the plurality of training samples include positive samples indicating recommendation and negative samples indicating recommendation;

[0020] A division unit, used for dividing the plurality of training samples into a training set and a test set;

[0021] The training unit is used to iteratively train the pre-built multi-objective model according to the training set, the test set and the label to obtain a refined multi-objective model for predicting the recommendation index of the video for a specific user.

[0022] According to a fifth aspect of the present disclosure, there is provided an electronic device, including:

[0023] at least one processor; and

[0024] a memory communicatively connected to the at least one processor; wherein,

[0025] The memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the method described in any one aspect.

[0026] According to a sixth aspect of the present disclosure, a non-transitory computer-readable storage medium storing computer instructions is provided, wherein the computer instructions are used to enable the computer to execute the method described in any one of the aspects.

[0027] According to a seventh aspect of the present disclosure, a computer program product is provided, the computer program product comprising: a computer program, the computer program being stored in a readable storage medium, at least one processor of an electronic device being able to read the computer program from the readable storage medium, and the at least one processor executing the computer program so that the electronic device executes the method described in any one aspect.

[0028] The technology disclosed herein solves the problem of how to recommend videos to users more flexibly and accurately and improve user experience.

[0029] 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

[0030] The accompanying drawings are used to better understand the present solution and do not constitute a limitation of the present disclosure.

[0031] Figure 1 It is a scene graph that can implement the video recommendation method and model training method of the embodiments of the present disclosure;

[0032] Figure 2 It is a system architecture diagram that can implement the video recommendation method and model training method of the embodiments of the present disclosure;

[0033] Figure 3 A flowchart of a video recommendation method provided in the first embodiment of the present disclosure;

[0034] Figure 4 A flowchart of a video recommendation method provided in the second embodiment of the present disclosure;

[0035] Figure 5 A schematic diagram of a flow chart of a video recommendation method provided in the third embodiment of the present disclosure;

[0036] Figure 6 A schematic diagram of a flow chart of a video recommendation method provided in the fourth embodiment of the present disclosure;

[0037] Figure 7 A flowchart of a model training method provided in the fifth embodiment of the present disclosure;

[0038] Figure 8A flowchart of a model training method provided for the sixth embodiment of the present disclosure;

[0039] Fig. 9 is a schematic diagram of an exemplary multi-objective model structure;

[0040] Fig.10 A structural block diagram of a video recommendation device provided by an embodiment of the present disclosure;

[0041] Fig.11 A structural block diagram of a model training device provided in an embodiment of the present disclosure;

[0042] Fig.12 A schematic block diagram of an electronic device for implementing an embodiment of the present disclosure is provided. DETAILED DESCRIPTION

[0043] The following is a description of exemplary embodiments of the present disclosure in conjunction with the accompanying drawings, including various details of the embodiments of the present disclosure to facilitate understanding, which should be considered as merely exemplary. Therefore, it should be recognized by those of ordinary skill in the art that various changes and modifications may be made to the embodiments described herein without departing from the scope and spirit of the present disclosure. Similarly, for the sake of clarity and conciseness, descriptions of well-known functions and structures are omitted in the following description.

[0044] In the related technologies, with the rapid development of mobile Internet technology, the penetration rate of smart phones has increased significantly, providing a solid foundation for the rise of short video applications. Users can watch, create and share short videos through their mobile phones anytime and anywhere, which greatly improves the content dissemination efficiency and user participation. In the fast-paced life, people are increasingly inclined to obtain entertainment and relaxation through short videos. Short videos meet users' fragmented entertainment needs with their short and concise features and rich content. As an emerging way of information dissemination, short videos convey information in an intuitive and vivid video form, which is easier for users to accept and understand. Users can interact with other users through likes, comments, sharing, etc. to form a social relationship network. Nowadays, most short video applications usually recommend videos to users by analyzing the content features of videos through filtering algorithms. The similarity between video features and user features is calculated, so that videos with high similarity are recommended to users. However, this method cannot fully consider the differences in video forms and user preference behaviors, and cannot accurately portray the user's habit of staying or swiping away.

[0045] In view of the existing problems, the technical conception process of the present disclosure is as follows: in the process of studying the recommendation of videos to users through applications, the inventors found that the application recommends videos to users through filtering algorithms, by analyzing the user's behavior data, finding other users with similar interests to the user, and then recommending the videos liked by these similar users to the current user; or by analyzing the similarity between videos, finding other videos similar to the videos that the user already likes, and then recommending these similar videos to the user. However, the interaction data between users and videos (such as viewing, liking, commenting, etc.) is sparse, that is, most users do not have clear ratings or feedback on most videos. This data sparsity will lead to insufficient accuracy in the representation of user interests, thereby affecting the recommendation effect. Therefore, the inventors consider whether it is possible to build a multi-objective model for fine sorting, and establish a user-granular personalized goal based on the user's stay time to characterize the user's habit of staying or swiping away, and then train the model. When the trained model is actually applied, based on the target user's stay time, distinguish the video form and select the closest target to take effect, thereby achieving more flexible and accurate video recommendation.

[0046] It should be noted that in the technical solution of the present disclosure, the collection, storage, use, processing, transmission, provision and disclosure of user personal information involved are in compliance with the provisions of relevant laws and regulations and do not violate public order and good morals.

[0047] Figure 1 is a scene graph that can implement the video recommendation method and model training method of the embodiment of the present disclosure, Figure 2 is a system architecture diagram that can implement the video recommendation method and model training method of the disclosed embodiment, combined with Figure 1 and Figure 2 As shown, the scene includes at least an electronic device 100, which can be a laptop, a smart phone, a desktop computer, a smart TV, a virtual reality (VR) device, a tablet computer, etc., which provides users with personalized video content and services through different technologies and methods to meet the user's diverse viewing needs. The electronic device 100 is configured with an application 101, a display screen 102, a memory 103 and a processor 104. Among them, the user can start and run the application 101 through the electronic device 100. After the application 101 is started, the processor 104 selects and roughly sorts multiple videos based on the user's historical viewing data in the memory 103 through a built-in algorithm, thereby obtaining a candidate video list, and then through the built-in trained multi-objective model for fine sorting, based on the user's stay time, based on different video types, determine the playback time label, so as to calculate the recommendation index based on the playback time label and other labels, sort based on the size of the recommendation index, obtain the fine sorting result, that is, the recommendation list, and display it to the user through the display screen 102.

[0048] Optionally, the electronic device 100 provides the functions of video recommendation and video viewing through the above-mentioned components. The communication mode between the above-mentioned components is usually realized through the bus system or specific interface protocol inside the electronic device. The application 101 is an interface for the user to interact with the electronic device 100, which is used to display video content, receive user input (such as search, click, etc.) and call the functions of other components. The application 101 is usually installed on the operating system of the electronic device 100 and interacts with the underlying hardware and software through the system API (application program interface). The display screen 102 is used to display video content and other visual information. The display screen 102 is usually connected to the processor 104 through a video output interface (such as HDMI, DisplayPort, LVDS, etc.), which can transmit high-quality video signals to ensure clear display of video content. The memory 103 is used to store video files, application data and various files of the operating system. The memory 103 can be an internal memory (such as RAM, ROM, flash memory, etc.) or an external memory (such as SD card, hard disk, etc.). The memory 103 communicates with the processor 104 via a memory bus or a specific storage interface (such as SATA, PCIe, etc.) to achieve fast data reading and writing. The processor 104 is the core component of the electronic device 100, responsible for executing program instructions, processing data, and controlling other components. The processor 104 communicates with each component through an internal bus system, including data exchange with the memory 103, video signal transmission with the display screen 102, and instruction execution with the application 101.

[0049] like Figure 2 As shown, before the application 101 is used, the processor 104 can extract features based on the user log, that is, the user's historical viewing data (stored in the memory 103), and combine the pre-set tags associated with the user's preferences (corresponding to Figure 2 After training, the trained model is configured in the processor 104 component of the electronic device 100. When the user starts the application 101, that is, in response to the user's request, the processor 104 performs multi-scale estimation and other estimations based on the video and user information through the trained model, that is, estimates the score of each tag, and then obtains the recommendation index corresponding to each video.

[0050] It should be noted that, in the above multi-scale estimation, it is necessary to determine a personalized playback duration label based on the user's historical consumption preference, that is, the user's stay time, so as to perform personalized estimation.

[0051] Understandably, Figure 1 and Figure 2The scene diagram shown is only an exemplary description. In actual applications, the scene diagram may also include other devices, such as network devices, audio devices, etc., which can be adjusted according to actual needs, and the present disclosure does not limit them. The embodiments of the present disclosure do not limit the actual forms of the various devices included in the application scenario, nor do they limit the interaction methods between the devices. In the specific application of the solution, they can be set according to actual needs.

[0052] The present disclosure provides a video recommendation method, a model training method, an apparatus, a device and a storage medium, which are applied to deep learning in the field of artificial intelligence technology, and establish user-granular personalized goals based on the quantiles of user stay time to characterize the user's habit of staying or swiping away, so as to achieve more flexible and accurate video recommendation for users and improve user experience.

[0053] The following specific embodiments are used to describe in detail the technical solution of the present invention and how the technical solution of the present invention solves the above-mentioned technical problems. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments. The embodiments of the present invention will be described below in conjunction with the accompanying drawings.

[0054] The specific implementation scheme of the video recommendation method provided by the present invention is introduced below.

[0055] Figure 3 A flowchart of a video recommendation method provided by the first embodiment of the present disclosure is shown in FIG. Figure 3 The video recommendation method shown in the figure can be configured as a video recommendation device, which can be located in an electronic device. The video recommendation may include the following steps:

[0056] S301: For each candidate video in the pre-acquired candidate video set, a pre-trained precision ranking multi-objective model is used to predict the candidate video according to the historical viewing data of the target user to obtain a recommendation index of the candidate video.

[0057] In this step, after the rough sorting process, the videos are screened for the target user to obtain the candidate video set. In order to make more flexible and accurate personalized recommendations for the user, the fine sorting multi-objective model is used to predict each candidate video based on the historical viewing data of the target user, thereby obtaining the recommendation index of each candidate video. The fine sorting multi-objective model is a model trained based on pre-set labels associated with user preferences and used to predict the recommendation index of a video for a specific user.

[0058] Specifically, the historical viewing data and each candidate video in the candidate video set are first preprocessed to obtain the target user information and the video information corresponding to each candidate video, and then based on the target user information, the preferred duration corresponding to the target user is determined, wherein the preferred duration includes the user preferred playback duration corresponding to at least one video type, and according to the video information, target user information and preferred duration corresponding to the candidate video, the scores of multiple labels corresponding to the candidate video are calculated through a refined multi-objective model to obtain the score corresponding to each label, and the score corresponding to all labels is combined to determine the recommendation index of the candidate video.

[0059] Exemplarily, historical viewing data may include viewing behavior data of target users, such as video ID or title, the time, progress, viewing frequency and dwell time of each video watched by the target user, and interactive behavior data, such as social interactive behavior data of the target user such as likes, shares and comments on each video, and data on the target user's interaction with the video during the viewing process, such as pause, fast forward, rewind and other operations, collection or playlist addition data, etc.

[0060] Optionally, it may also include metadata of videos watched by the target user, such as video title, description, tag, category, release time, author or creator information, etc. It may also include video quality indicators, such as video resolution, bit rate, frame rate, etc. It may also include video content features, such as scenes, characters, emotions, topics, etc.

[0061] In one possible implementation, the pre-set tags associated with user preferences include at least two playback duration tags, each playback duration tag represents the user's dwell time on the video preference, and may also include any one or more of a satisfaction tag, a completion rate tag, a resource debiasing tag, a mid-play tag, a finished play tag, and a fast-scroll tag.

[0062] The satisfaction label is the video playback time in the physical time bucket. The top 20% of the buckets are recommended positive samples, and those less than 10 seconds are not recommended negative samples.

[0063] The completion rate label is video playback time / video resource time.

[0064] Resource debiasing labels are used to debias the actual playback duration by bucketing it according to the physical duration, and the thousandths in the bucket after debiasing are used as labels.

[0065] For low-activity users, the medium-play label is a recommended positive sample if the playback time is longer than 30 seconds. For medium-activity users, the playback time is longer than 40 seconds. For high-activity users, the playback time is longer than 50 seconds. It is a positive sample, otherwise it is a negative sample.

[0066] The completion label is a positive sample when the completion rate is 95%, otherwise it is a negative sample.

[0067] The fast scrolling label is a positive sample when the playback time is less than 3s, otherwise it is a negative sample.

[0068] It should be noted that the above setting of each tag is only an example, and the user can pre-set it according to the time scenario, and this embodiment does not make any specific limitation.

[0069] S302: Sort the candidate videos in the candidate video set from large to small according to the recommendation index to obtain a recommendation list.

[0070] S303: Push the recommendation list to the target user.

[0071] In this step, after predicting the recommendation index of each candidate video through the above steps, all candidate videos are sorted based on the size of the recommendation index to obtain a recommendation list, and the recommendation list is pushed to the target user.

[0072] For example, when sorting candidate videos, they are sorted from large to small or from small to large according to the recommendation index. The specific sorting method is not specifically limited. After obtaining the recommendation list, the recommendation category includes the video content that best meets the interests and preferences of the target user, and is displayed in a certain format (such as a list, a grid, etc.). The generated recommendation list is pushed to the target user through appropriate channels (such as application push, web page display, email notification, etc.). After receiving the recommendation list, the target user can choose to watch the corresponding video content according to his or her interests and needs.

[0073] Optionally, the recommendation list may include all candidate videos, or may include any one or more of all candidate videos, and may be actually set according to actual recommendation scenarios, and is not specifically limited here.

[0074] The video recommendation method provided in this embodiment predicts each candidate video in the pre-acquired candidate video set based on the historical viewing data of the target user using a pre-trained precision ranking multi-objective model to obtain the recommendation index of the candidate video, and sorts the candidate videos in the candidate video set according to the size of the recommendation index to obtain a recommendation list, which is then pushed to the target user. By considering the personalized data of the target user, the user's interest preferences and behavior patterns can be deeply understood. The precision ranking multi-objective model uses the personalized information of the target user to generate customized recommendation content for each target user, thereby significantly improving the accuracy and personalization of the recommendation, and reducing the operational complexity of the user's query retrieval of preferred videos by providing personalized, high-quality recommendation content.

[0075] Figure 4 A flowchart of a video recommendation method provided in the second embodiment of the present disclosure is shown in FIG. Figure 4 The video recommendation method shown is different from the aforementioned embodiment in that, in the present embodiment, a pre-trained precision ranking multi-objective model is used to predict candidate videos based on the historical viewing data of the target user to obtain the recommendation index of the candidate video, which can be refined as follows: pre-processing the historical viewing data and each candidate video in the candidate video set to obtain the target user information and the video information corresponding to each candidate video, determining the preferred duration corresponding to the target user based on the target user information, and for each candidate video, calculating the scores of multiple tags corresponding to the candidate video through the precision ranking multi-objective model based on the video information, target user information and preferred duration corresponding to the candidate video to obtain the score corresponding to each tag, and determining the recommendation index of the candidate video based on the scores corresponding to all tags.

[0076] like Figure 4 As shown, the method comprises the following steps:

[0077] S401: Preprocess the historical viewing data and each candidate video in the candidate video set to obtain target user information and video information corresponding to each candidate video.

[0078] In this step, in order to recommend videos that are more in line with the target user's personalized preferences, it is necessary to pre-process the target user's historical viewing data and each candidate video in the candidate video set.

[0079] Specifically, the historical viewing data of the target user includes the historical viewing behavior data of the target user, the interaction data with the watched videos, etc. In order to extract the features that reflect the user's interests and preferences, data cleaning can be performed first to remove duplicate, erroneous or invalid records, such as abnormal viewing time, missing video IDs, etc. Then feature extraction is performed to extract user behavior features, such as viewing time, viewing frequency, whether to like, share, comment, etc. Video preference features are extracted, such as extracting preference tags based on the type of videos watched by the target user, themes, actors, directors, etc. Time features are extracted, such as viewing time (such as weekdays, weekends, daytime, night), viewing time, dwell time, etc.

[0080] For each candidate video, information is extracted during preprocessing, such as video metadata extraction, such as channel ID, title, description, tag, category, release time, author, duration, etc. Content feature extraction, such as scene, character, emotion, topic, etc., extracted through video analysis technology (such as image recognition, speech recognition, natural language processing, etc.) Quality indicators are extracted, such as video resolution, bit rate, frame rate, etc.

[0081] S402: Determine the preferred duration corresponding to the target user based on the target user information.

[0082] In this step, it can be seen from the above embodiments that the resource debiasing label in the set label takes into account the debiasing of the resource form, that is, the video type, but does not take into account the differences in user consumption behaviors and cannot accurately describe the user's habit of staying or swiping away. In order to differentiate according to the user's consumption history and improve the recommendation experience, the preferred duration corresponding to the target user is determined based on the target user information.

[0083] Exemplarily, when training the multi-objective model for precise ranking, it is trained based on labels associated with user preferences. The labels include at least two playback time labels, namely multi-scale playback time. Each playback time represents the user's preferred dwell time on the video. For example, based on the statistics of the general population's painting style, four playback time labels of 10s, 20s, 30s and 50s can be set. Based on the target user information, the user's preferred duration for different types of videos is analyzed to determine that the target user is accustomed to consuming small videos of about 10s and short videos of about 30s.

[0084] S403: For each candidate video, according to the video information, target user information and preferred duration corresponding to the candidate video, the scores of multiple tags corresponding to the candidate video are calculated by a refined multi-objective model to obtain a score corresponding to each tag.

[0085] In this step, after obtaining the video information, target user information and the preferred duration corresponding to each candidate video through the above steps, the multiple label scores corresponding to each candidate video are calculated through the refined multi-objective model.

[0086] Specifically, for each candidate video, the video type of the candidate video is determined based on the video information corresponding to the candidate video, and the score of each playback time label corresponding to the candidate video is calculated through the precision sorting multi-objective model based on the video information corresponding to the candidate video, the target user information and the user preferred playback time corresponding to the video type. The scores of other labels are calculated through the precision sorting multi-objective model based on the video information corresponding to the candidate video and the target user information. Finally, the score corresponding to each label is obtained.

[0087] S404: Determine the recommendation index of the candidate video based on the scores corresponding to all tags.

[0088] Exemplarily, after obtaining the scores corresponding to all tags, the weight corresponding to each tag can be determined, and based on the weight corresponding to each tag, the scores corresponding to all tags are weightedly fused and calculated to obtain the recommendation index of the candidate video.

[0089] S405: Sort the candidate videos in the candidate video set from large to small according to the recommendation index to obtain a recommendation list.

[0090] S406: Push the recommendation list to the target user.

[0091] The specific implementation of step S405 and step S406 is the same as that in the above-mentioned embodiment, and will not be repeated here.

[0092] The video recommendation method provided in this embodiment pre-processes the historical viewing data and each candidate video in the candidate video set to obtain the target user information and the video information corresponding to each candidate video, and determines the preferred duration corresponding to the target user based on the target user information. For each candidate video, the scores of multiple tags corresponding to the candidate video are calculated through a precision sorting multi-objective model according to the video information, target user information and preferred duration corresponding to the candidate video to obtain the score corresponding to each tag, and the recommendation index of the candidate video is determined based on the scores corresponding to all tags, and the candidate videos in the candidate video set are sorted according to the size of the recommendation index to obtain a recommendation list, which is pushed to the target user. This method improves the flexibility and accuracy of video recommendation through personalized recommendation, multi-factor consideration, multi-objective optimization, efficient sorting, and dynamic adjustment.

[0093] Figure 5 A flowchart of a video recommendation method provided in the third embodiment of the present disclosure is shown in FIG. Figure 5 The video recommendation method shown is different from the aforementioned embodiment in that, in the present embodiment, multiple label scores corresponding to the candidate videos are calculated by a refined ranking multi-objective model based on the video information, target user information and preferred duration corresponding to the candidate videos. It can also be refined as follows: for each candidate video, the video type of the candidate video is determined based on the video information corresponding to the candidate video, and based on the video information, target user information and user preferred playback duration corresponding to the video type corresponding to the candidate video, each playback duration label score corresponding to the candidate video is calculated by a refined ranking multi-objective model; and other label scores are calculated by a refined ranking multi-objective model based on the video information and target user information corresponding to the candidate video.

[0094] like Figure 5 As shown, the method comprises the following steps:

[0095] S501: Preprocess the historical viewing data and each candidate video in the candidate video set to obtain target user information and video information corresponding to each candidate video.

[0096] S502: Determine the preferred duration corresponding to the target user based on the target user information.

[0097] The specific implementation of step S501 and step S502 is the same as that in the above-mentioned embodiment, and will not be repeated here.

[0098] S503: For each candidate video, determine the video type of the candidate video according to the video information corresponding to the candidate video.

[0099] In this step, in order to finely characterize the user's habit of staying or swiping away, at least two playback duration labels are set for training based on the user's stay time during the multi-objective model training process. When the model is applied, in order to reflect the differences in consumption habits of different users, the target user's corresponding preferred duration is determined based on the target user information. Then, when the model is predicted, the calculation of the playback duration label score only calculates the label score corresponding to the determined preferred duration.

[0100] Specifically, the video forms are distinguished to determine the target user's preferred playback duration corresponding to different video forms, and then the type of each candidate video is determined based on the video information corresponding to the candidate video.

[0101] Exemplarily, the video information may be extracted, and then the actual length of the video may be obtained by parsing the extracted video file or using an API, and then the video may be classified into short videos, small videos, or other types according to a preset threshold.

[0102] It should be noted that the above-mentioned types of small videos and short videos are merely examples, and the specific types of classification can be set according to actual conditions, and this embodiment does not make any specific limitations.

[0103] S504: Based on the video information corresponding to the candidate video, the target user information and the user preferred playback time corresponding to the video type, the score of each playback time label corresponding to the candidate video is calculated through the refined multi-objective model.

[0104] In this step, after determining the type of each candidate video, for each candidate video, based on the video information corresponding to the candidate video, the target user information and the user preferred playback time corresponding to the video type, the precise sorting multi-objective model is used to calculate the label score of each playback time corresponding to the candidate video.

[0105] For example, if the target user is used to consuming small videos of about 10 seconds and short videos of about 30 seconds, the 10-second playback target and the 30-second playback target are respectively implemented according to the video form. That is, for short videos, the 30-second playback time label score is calculated, and for small videos, the 10-second playback time label score is calculated.

[0106] S505: Calculate other label scores through a refined ranking multi-objective model according to the video information and target user information corresponding to the candidate video.

[0107] In this step, the calculation method for other label scores is the same as the above calculation method, which will not be repeated here.

[0108] S506: Determine the recommendation index of the candidate video based on the scores corresponding to all tags.

[0109] S507: Sort the candidate videos in the candidate video set from large to small according to the recommendation index to obtain a recommendation list.

[0110] S508: Push the recommendation list to the target user.

[0111] The specific implementation of step S506 to step S508 is the same as that in the above-mentioned embodiment, and will not be repeated here.

[0112] The video recommendation method provided in this embodiment pre-processes each candidate video in the historical viewing data and the candidate video set, obtains the target user information and the video information corresponding to each candidate video, determines the target user's corresponding preferred duration based on the target user information, determines the video type of the candidate video based on the video information corresponding to the candidate video, calculates the score of each playback duration label corresponding to the candidate video through the precision sorting multi-objective model based on the video information corresponding to the candidate video, the target user information and the user preferred playback duration corresponding to the video type, calculates the scores of other labels through the precision sorting multi-objective model based on the video information corresponding to the candidate video and the target user information, determines the recommendation index of the candidate video based on the scores corresponding to all labels, sorts the candidate videos in the candidate video set according to the size of the recommendation index, obtains a recommendation list, and pushes the recommendation list to the target user. Through the above method, since the recommendation is more personalized and accurate, it is easier for users to find their favorite videos, thereby reducing the operational complexity of the user's query retrieval of the preferred video. By optimizing the playback duration and other interactive indicators, the accuracy and personalization of the recommendation are improved.

[0113] Figure 6 A flowchart of a video recommendation method provided in the fourth embodiment of the present disclosure is shown in FIG. Figure 6 The video recommendation method shown is different from the aforementioned embodiment in that, in this embodiment, the recommendation index of the candidate video is determined based on the scores corresponding to all tags, and can be further refined as follows: for each tag, the weight corresponding to the tag is determined, and based on the weight corresponding to each tag, the scores corresponding to all tags are weightedly fused and calculated to obtain the recommendation index of the candidate video.

[0114] like Figure 6As shown, the method comprises the following steps:

[0115] S601: Preprocess the historical viewing data and each candidate video in the candidate video set to obtain target user information and video information corresponding to each candidate video.

[0116] S602: Determine the preferred duration corresponding to the target user based on the target user information.

[0117] S603: For each candidate video, according to the video information, target user information and preferred duration corresponding to the candidate video, the scores of multiple tags corresponding to the candidate video are calculated by a refined multi-objective model to obtain a score corresponding to each tag.

[0118] The specific implementation of step S601 to step S603 is the same as that in the above-mentioned embodiment, and will not be repeated here.

[0119] S604: For each tag, determine the weight corresponding to the tag.

[0120] S605: Based on the weight corresponding to each tag, a weighted fusion calculation is performed on the scores corresponding to all tags to obtain a recommendation index of the candidate video.

[0121] In order to make personalized video recommendations for users more flexibly and accurately, after obtaining the score of each tag, a weight can be set for each tag score, and then through weighted fusion, the recommendation index of each candidate video can be obtained.

[0122] For example, the weights can be determined by assigning fixed weights to each tag based on business needs and experience, or by counting the behavior probabilities of target users under different tags and then calculating weights for each tag according to importance. The weights of tags can also be automatically learned through machine learning models, or the tag weights can be dynamically adjusted based on system performance and user feedback.

[0123] S606: Sort the candidate videos in the candidate video set from large to small according to the recommendation index to obtain a recommendation list.

[0124] S607: Push the recommendation list to the target user.

[0125] The specific implementation of step S606 and step S607 is the same as that in the above-mentioned embodiment, and will not be repeated here.

[0126] The video recommendation method provided in this embodiment pre-processes each candidate video in the historical viewing data and the candidate video set, obtains the target user information and the video information corresponding to each candidate video, determines the preferred duration corresponding to the target user based on the target user information, and for each candidate video, calculates the scores of multiple tags corresponding to the candidate video through the precision sorting multi-objective model according to the video information, target user information and preferred duration corresponding to the candidate video, obtains the score corresponding to each tag, determines the weight corresponding to the tag for each tag, and performs weighted fusion calculation on the scores corresponding to all tags based on the weight corresponding to each tag, obtains the recommendation index of the candidate video, sorts the candidate videos in the candidate video set according to the size of the recommendation index, obtains the recommendation list, and pushes the recommendation list to the target user. The precision sorting multi-objective model is used to calculate the scores of multiple tags of the candidate video, through considerations of multiple dimensions. The conflicts between different goals are balanced, the overall performance of the recommendation system is improved, and the accuracy and flexibility of video recommendation are improved by pushing video content that meets the interests and preferences of users.

[0127] The specific implementation scheme of the model training method provided by the present invention is introduced below.

[0128] Figure 7 A flowchart of a model training method provided in the fifth embodiment of the present disclosure is provided. Figure 7 The model training method shown in the figure specifically comprises the following steps:

[0129] S701: Calibrate historical viewing data corresponding to multiple users based on tags associated with user preferences to obtain multiple training samples.

[0130] In this step, in order to be able to make video recommendations for users more flexibly and accurately, and to be able to differentiate based on user consumption history and improve the recommendation experience, labels associated with user preferences are pre-set, and then based on the labels, the historical viewing data corresponding to multiple users are calibrated to obtain multiple training samples, wherein the multiple training samples include positive samples indicating recommendations and negative samples indicating recommendations.

[0131] Exemplarily, the label includes at least two playback time labels, each playback time label represents the user's preferred dwell time on the video. For example, based on the percentile of the user's dwell time, a user-granular personalized goal is established to characterize the user's habit of staying or swiping away. According to the statistics of the general population's painting style, four levels of playback time labels are distinguished, namely 10s, 20s, 30s and 50s.

[0132] Optionally, the label may further include any one or more of a satisfaction label, a completion rate label, a resource debiasing label, a mid-broadcast label, a completed broadcast label, and a fast-scroll label.

[0133] The satisfaction label is the video playback time in the physical time bucket. The first 20% of the buckets are positive samples indicating recommendation, and those less than 10s are negative samples indicating recommendation.

[0134] The completion rate label is video playback time / video resource time.

[0135] Resource debiasing labels are used to debias the actual playback duration by bucketing it according to the physical duration, and the thousandths in the bucket after debiasing are used as labels.

[0136] For low-activity users, the medium-playing label indicates a positive sample for recommendation if the playback time is longer than 30 seconds. For medium-activity users, the medium-playing label indicates a positive sample for recommendation if the playback time is longer than 40 seconds. For high-activity users, the medium-playing label indicates a positive sample for recommendation if the playback time is longer than 50 seconds, otherwise it indicates a negative sample for recommendation.

[0137] The completion label indicates that a positive sample is recommended when the completion rate reaches 95%, otherwise a negative sample is recommended.

[0138] The fast scrolling label indicates that a positive sample is recommended if the playback time is less than 3 seconds, otherwise it indicates that a negative sample is recommended.

[0139] The historical viewing data corresponding to multiple users are calibrated based on the above tags.

[0140] It should be noted that the above setting of each tag is only an example, and the user can pre-set it according to the time scenario, and this embodiment does not make any specific limitation.

[0141] S702: Divide the multiple training samples into a training set and a test set.

[0142] In this step, after a plurality of training samples are obtained, the plurality of training samples may be divided into a training set and a test set.

[0143] Exemplarily, the training samples may be divided in half, with 50% of the training samples used as training sets and 50% used as test sets, or 70% of the training samples may be used as training sets and 20% used as test sets. This embodiment does not impose any specific restrictions on the specific division ratio.

[0144] S703: Iteratively train the pre-built multi-objective model according to the training set, the test set and the labels to obtain a refined multi-objective model for predicting the recommendation index of the video for a specific user.

[0145] In this step, after the training set and the test set are divided, the pre-built multi-objective model is iteratively trained based on the training set until the model converges, and then verification and optimization are performed based on the test set to obtain a refined multi-objective model for predicting the recommendation index of a video for a specific user.

[0146] Specifically, according to the training set and the test set, based on each label, the tower corresponding to each label is iteratively trained to obtain a refined multi-objective model.

[0147] It should be noted that the labels serve as supervisory signals during the training process to supervise the training process.

[0148] The model training method provided in this embodiment calibrates the historical viewing data corresponding to multiple users based on the labels associated with the user preferences to obtain multiple training samples, divides the multiple training samples into a training set and a test set, and iteratively trains the pre-built multi-objective model according to the training set, the test set and the labels to obtain a refined multi-objective model for predicting the recommendation index of a video for a specific user. By calibrating the historical viewing data using labels that are closely related to the user's preferences, the unique interests and preferences of each user can be captured more accurately. The model can learn the personalized characteristics of the user during the training process, thereby recommending video content that is more in line with the user's personal preferences, thereby improving the accuracy and personalization of the recommendation. Using a multi-objective model for training can simultaneously consider multiple evaluation indicators and balance the conflicts between different business goals.

[0149] Figure 8 A flow chart of a model training method provided in the sixth embodiment of the present disclosure, referring to Figure 8 The model training method shown is different from the aforementioned embodiment in that, in the present embodiment, a multi-objective model can also be constructed. The multi-objective model includes a shared underlying network architecture and multiple towers. Accordingly, the multi-objective model is iteratively trained according to the training set, the test set and the labels to obtain a refined multi-objective model for predicting the recommendation index of a video for a specific user. The refined multi-objective model can be further refined as follows: according to the training set and the test set, based on each label, the tower corresponding to each label is iteratively trained to obtain a refined multi-objective model.

[0150] like Figure 8 As shown, the method comprises the following steps:

[0151] S801: Calibrate historical viewing data corresponding to multiple users based on tags associated with user preferences to obtain multiple training samples.

[0152] S802: Divide multiple training samples into a training set and a test set.

[0153] The specific implementation of step S801 and step S802 is the same as that in the above-mentioned embodiment, and will not be repeated here.

[0154] S803: Construct a multi-objective model.

[0155] In this step, in order to recommend videos to users more flexibly and accurately, a multi-objective model is pre-built and trained offline to obtain a refined multi-objective model for predicting the recommendation index of a video for a specific user.

[0156] Exemplarily, the constructed multi-objective model structure may include a shared underlying network architecture and multiple towers. Fig. 9 is a schematic diagram of an exemplary multi-objective model structure, such as Fig. 9 As shown in the figure, the preset playback time labels are 10s, 20s, 30s and 50s, a total of 4 levels. Fig. 9 In the example, level 1, level 2, level 3, and level 4 correspond to different playback time labels, and each label corresponds to an independent tower, sharing the Expert, thereby utilizing information or experts of different scales to process data. In addition, the underlying network architecture is shared to ensure information sharing between different tasks at the feature level. In multi-task learning, since multiple tasks share the underlying network, the learning process of the underlying network will be affected by the loss functions of multiple tasks. This joint training method helps to reduce the risk of overfitting. By sharing the underlying network, the model can learn more generalized feature representations. These features are not only applicable to the current set of tasks, but may also have certain generalization capabilities for unseen tasks. In some cases, sharing the underlying network can also help the model implicitly model the relationship between different tasks. Although this relationship may not be directly defined by the model structure, through the learning process of sharing the underlying network, the model can learn the potential connection between different tasks.

[0157] S804: According to the training set and the test set, based on each label, iteratively train the tower corresponding to each label to obtain a refined multi-objective model.

[0158] In this step, after multiple labels are preset, since different targets correspond to independent towers, the corresponding towers can be iteratively trained based on different labels. During the joint training process, multiple independent towers share the underlying network, thereby sharing information and making the model more accurate.

[0159] The model training method provided in this embodiment calibrates the historical viewing data corresponding to multiple users based on the labels associated with the user preferences, obtains multiple training samples, divides the multiple training samples into training sets and test sets, constructs a multi-objective model, and iteratively trains the tower corresponding to each label based on the training set and the test set to obtain a refined multi-objective model. By associating the label with the user's historical data, the user's interests and preferences can be understood more accurately, thereby improving the relevance of the recommendation. The training towers corresponding to different labels can consider multiple objectives at the same time, so that the model is optimized in multiple dimensions, rather than focusing on a single objective, thereby improving the quality of recommendations. Through the multi-objective model and the refined ranking process, the generalization ability of the model can be enhanced, and the flexibility of the model can also be guaranteed.

[0160] Fig.10 A structural block diagram of a video recommendation device provided by an embodiment of the present disclosure, referring to Fig.10 The video recommendation device 1000 shown includes the following units:

[0161] The prediction unit 1001 is used to predict each candidate video in the pre-acquired candidate video set according to the historical viewing data of the target user using a pre-trained precision ranking multi-objective model to obtain a recommendation index of the candidate video, where the precision ranking multi-objective model is a model trained based on pre-set labels associated with user preferences and used to predict the recommendation index of a video for a specific user;

[0162] A sorting unit 1002 is used to sort the candidate videos in the candidate video set from large to small according to the recommendation index to obtain a recommendation list;

[0163] The push unit 1003 is used to push the recommendation list to the target user.

[0164] In a possible design, the tag includes at least two playback duration tags, each of which indicates the duration of a user's preference for a video.

[0165] Optionally, the tag also includes at least one of the following tags:

[0166] Satisfaction label, completion rate label, resource debiasing label, mid-broadcast label, finished broadcast label, and fast-swipe label.

[0167] In another possible design, the prediction unit 1001 includes:

[0168] A preprocessing module, used to preprocess the historical viewing data and each candidate video in the candidate video set to obtain target user information and video information corresponding to each candidate video;

[0169] A first determination module is used to determine a preferred duration corresponding to the target user based on the target user information, where the preferred duration includes a user preferred playback duration corresponding to at least one video type;

[0170] A calculation module is used to calculate the scores of multiple tags corresponding to each candidate video through a refined multi-objective model according to the video information, target user information and preferred duration corresponding to the candidate video, so as to obtain the score corresponding to each tag;

[0171] The second determination module is used to determine the recommendation index of the candidate video based on the scores corresponding to all the tags.

[0172] Optional, computing module, including:

[0173] A determination submodule, used to determine the video type of each candidate video according to the video information corresponding to the candidate video;

[0174] The first calculation submodule is used to calculate the score of each playback time label corresponding to the candidate video through the refined multi-objective model based on the video information corresponding to the candidate video, the target user information and the user preferred playback time corresponding to the video type;

[0175] The second calculation submodule is used to calculate the scores of other tags through the refined ranking multi-objective model according to the video information and target user information corresponding to the candidate video.

[0176] Optionally, the second determination module includes:

[0177] A determination submodule is used to determine the weight corresponding to each tag;

[0178] The calculation submodule is used to perform weighted fusion calculation on the scores corresponding to all tags based on the weight corresponding to each tag, so as to obtain the recommendation index of the candidate video.

[0179] Fig.11 A structural block diagram of a model training device provided in an embodiment of the present disclosure, refer to Fig.11 The model training device 1100 shown includes the following units:

[0180] The calibration unit 1101 is used to calibrate the historical viewing data corresponding to multiple users based on the tags associated with the user preferences to obtain multiple training samples, wherein the multiple training samples include positive samples indicating recommendation and negative samples indicating recommendation;

[0181] A division unit 1102 is used to divide a plurality of training samples into a training set and a test set;

[0182] The training unit 1103 is used to iteratively train the pre-built multi-objective model according to the training set, the test set and the labels to obtain a refined multi-objective model for predicting the recommendation index of the video for a specific user.

[0183] Optionally, the tag includes at least two playback duration tags, each playback duration tag indicating the duration of a user's preference for a video.

[0184] In a possible implementation, the tag further includes at least one of the following tags:

[0185] Satisfaction label, completion rate label, resource debiasing label, mid-broadcast label, finished broadcast label, and fast-swipe label.

[0186] In a possible implementation, the model training device 1100 further includes:

[0187] A construction unit 1104 is used to construct a multi-objective model, where the multi-objective model includes a shared underlying network architecture and multiple towers;

[0188] Among them, each label corresponds to a tower.

[0189] In another possible implementation, the training unit 1103 includes:

[0190] The training module is used to iteratively train the tower corresponding to each label based on the training set and the test set to obtain a refined multi-objective model.

[0191] According to an embodiment of the present disclosure, the present disclosure also provides an electronic device and a readable storage medium.

[0192] According to an embodiment of the present disclosure, the present disclosure also provides a computer program product, which includes: a computer program, the computer program is stored in a readable storage medium, at least one processor of an electronic device can read the computer program from the readable storage medium, and at least one processor executes the computer program so that the electronic device executes the solution provided by any of the above embodiments.

[0193] Fig.12 A schematic block diagram of an electronic device implementing an embodiment of the present disclosure is shown in FIG. Fig.12 As shown, Fig.12A schematic block diagram of an example electronic device 1200 that can be used to implement an embodiment of the present disclosure is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device can also represent various forms of mobile devices, such as personal digital processing, cellular phones, smart phones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely examples and are not intended to limit the implementation of the present disclosure described and / or required herein.

[0194] like Fig.12 As shown, the electronic device 1200 includes a computing unit 1201, which can perform various appropriate actions and processes according to a computer program stored in a read-only memory (ROM) 1202 or a computer program loaded from a storage unit into a random access memory (RAM) 1203. In the RAM 1203, various programs and data required for the operation of the electronic device 1200 can also be stored. The computing unit 1201, the ROM 1202, and the RAM 1203 are connected to each other via a bus 1204. An input / output (I / O) interface 1205 is also connected to the bus 1204.

[0195] Multiple components in the electronic device 1200 are connected to the I / O interface 1205, including: an input unit 1206, such as a keyboard, a mouse, etc.; an output unit 1207, such as various types of displays, speakers, etc.; a storage unit 1208, such as a disk, an optical disk, etc.; and a communication unit 1209, such as a network card, a modem, a wireless communication transceiver, etc. The communication unit 1209 allows the electronic device 1200 to exchange information / data with other devices through a computer network such as the Internet and / or various telecommunication networks.

[0196] The computing unit 1201 may be a variety of general and / or special processing components with processing and computing capabilities. Some examples of the computing unit 1201 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various dedicated artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, digital signal processors (DSPs), and any appropriate processors, controllers, microcontrollers, etc. The computing unit 1201 performs the various methods and processes described above, such as the method video recommendation method and the model training method. For example, in some embodiments, the method video recommendation method and the model training method may be implemented as a computer software program, which is tangibly included in a machine-readable medium, such as a storage unit 1208. In some embodiments, part or all of the computer program may be loaded and / or installed on the electronic device 1200 via the ROM 1202 and / or the communication unit 1209. When the computer program is loaded into the RAM 1203 and executed by the computing unit 1201, one or more steps of the method video recommendation method and the model training method described above may be executed. Alternatively, in other embodiments, the computing unit 1201 may be configured to execute the video recommendation method and the model training method in any other appropriate manner (eg, by means of firmware).

[0197] Various implementations of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field programmable gate arrays (FPGAs), application specific integrated circuits (ASICs), application specific standard products (ASSPs), systems on chips (SOCs), complex programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various implementations can include: being implemented in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which can be a special purpose or general purpose programmable processor that can receive data and instructions from a storage system, at least one input device, and at least one output device, and transmit data and instructions to the storage system, the at least one input device, and the at least one output device.

[0198] The program code for implementing the method of the present disclosure may be written in any combination of one or more programming languages. These program codes may be provided to a processor or controller of a general-purpose computer, a special-purpose computer, or other programmable data processing device, so that the program code, when executed by the processor or controller, enables the functions / operations specified in the flow chart and / or block diagram to be implemented. The program code may be executed entirely on the machine, partially on the machine, partially on the machine and partially on a remote machine as a stand-alone software package, or entirely on a remote machine or server.

[0199] In the context of the present disclosure, a machine-readable medium may be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, device, or equipment. A machine-readable medium may be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium may include, but is not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, device, or equipment, or any suitable combination of the foregoing. A more specific example of a machine-readable storage medium may include an electrical connection based on one or more lines, a portable computer 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 foregoing.

[0200] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to the user; and a keyboard and pointing device (e.g., a mouse or trackball) through which the user can provide input to the computer. Other types of devices can also be used to provide interaction with the user; for example, the feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including acoustic input, voice input, or tactile input).

[0201] The systems and techniques described herein may be implemented in a computing system that includes back-end components (e.g., as a data server), or a computing system that includes middleware components (e.g., an application server), or a computing system that includes front-end components (e.g., a user computer with a graphical user interface or a web browser through which a user can interact with implementations of the systems and techniques described herein), or a computing system that includes any combination of such back-end components, middleware components, or front-end components. The components of the system may be interconnected by any form or medium of digital data communication (e.g., a communication network). Examples of communication networks include: a local area network (LAN), a wide area network (WAN), and the Internet.

[0202] A computer system may include a client and a server. The client and the server are generally remote from each other and usually interact through a communication network. The relationship between the client and the server is generated by computer programs running on the corresponding computers and having a client-server relationship with each other. The server may be a cloud server, also known as a cloud computing server or cloud host, which is a host product in the cloud computing service system to solve the defects of difficult management and weak business scalability in traditional physical hosts and VPS services ("Virtual Private Server", or "VPS" for short). The server may also be a server of a distributed system, or a server combined with a blockchain.

[0203] It should be understood that the various forms of processes shown above can be used to reorder, add or delete steps. For example, the steps recorded in this disclosure can be executed in parallel, sequentially or in different orders, as long as the desired results of the technical solutions disclosed in this disclosure can be achieved, and this document does not limit this.

[0204] The above specific implementations do not constitute a limitation on the protection scope of the present disclosure. It should be understood by those skilled in the art that various modifications, combinations, sub-combinations and substitutions can be made according to design requirements and other factors. Any modification, equivalent substitution and improvement made within the spirit and principle of the present disclosure shall be included in the protection scope of the present disclosure.

Claims

1. A video recommendation method, comprising: For each candidate video in the pre-acquired candidate video set, a pre-trained precision ranking multi-objective model is used to predict the candidate video according to the historical viewing data of the target user to obtain a recommendation index of the candidate video, wherein the precision ranking multi-objective model is a model trained based on pre-set labels associated with user preferences and used to predict the recommendation index of a video for a specific user; Sorting the candidate videos in the candidate video set from large to small according to the recommendation index to obtain a recommendation list; The recommendation list is pushed to the target user.

2. The method according to claim 1, wherein: The tags include at least two play time tags, each of which indicates the duration of a user's preference for a video.

3. The method according to claim 2, wherein: The method of predicting the candidate videos based on the historical viewing data of the target user using a pre-trained precision ranking multi-objective model to obtain the recommendation index of the candidate videos includes: Preprocessing the historical viewing data and each candidate video in the candidate video set to obtain target user information and video information corresponding to each candidate video; Based on the target user information, determining a preferred duration corresponding to the target user, wherein the preferred duration includes a user preferred playback duration corresponding to at least one video type; For each candidate video, according to the video information corresponding to the candidate video, the target user information and the preferred duration, the multiple label scores corresponding to the candidate video are calculated by the refined multi-objective model to obtain a score corresponding to each label; Based on the scores corresponding to all tags, the recommendation index of the candidate video is determined.

4. The method according to claim 3, wherein: The calculating of multiple label scores corresponding to the candidate videos by the refined sorting multi-objective model according to the video information corresponding to the candidate videos, the target user information and the preferred duration includes: For each candidate video, determining the video type of the candidate video according to the video information corresponding to the candidate video; Based on the video information corresponding to the candidate video, the target user information and the user preferred playback time corresponding to the video type, the score of each playback time label corresponding to the candidate video is calculated by the refined multi-objective model; According to the video information corresponding to the candidate video and the target user information, the scores of other tags are calculated through the refined ranking multi-objective model.

5. The method according to claim 3, wherein: The step of determining the recommendation index of the candidate video based on the scores corresponding to all the tags includes: For each tag, determine the weight corresponding to the tag; Based on the weight corresponding to each tag, the scores corresponding to all tags are weighted and fused to obtain the recommendation index of the candidate video.

6. The method according to claim 2, wherein the tag further comprises at least one of the following tags: Satisfaction label, completion rate label, resource debiasing label, mid-broadcast label, finished broadcast label, and fast-swipe label.

7. A model training method, comprising: Calibrate historical viewing data corresponding to a plurality of users based on tags associated with user preferences to obtain a plurality of training samples, wherein the plurality of training samples include positive samples indicating recommendation and negative samples indicating recommendation; Dividing the multiple training samples into a training set and a test set; The pre-built multi-objective model is iteratively trained according to the training set, the test set and the labels to obtain a refined multi-objective model for predicting the recommendation index of a video for a specific user.

8. The method according to claim 7, wherein: The tags include at least two play time tags, each of which indicates the duration of a user's preference for a video.

9. The method according to claim 7, further comprising: Constructing the multi-objective model, wherein the multi-objective model includes a shared underlying network architecture and a plurality of towers; Among them, each label corresponds to a tower.

10. The method according to claim 9, wherein: The method of iteratively training the pre-built multi-objective model according to the training set, the test set and the label to obtain a refined multi-objective model for predicting the recommendation index of the video for a specific user includes: According to the training set and the test set, based on each label, the tower corresponding to each label is iteratively trained to obtain the refined multi-objective model.

11. The method according to claim 8, wherein the tag further comprises at least one of the following tags: Satisfaction label, completion rate label, resource debiasing label, mid-broadcast label, finished broadcast label, and fast-swipe label.

12. A video recommendation device, comprising: A prediction unit is used to predict each candidate video in the pre-acquired candidate video set according to the historical viewing data of the target user using a pre-trained precision ranking multi-objective model to obtain a recommendation index of the candidate video, wherein the precision ranking multi-objective model is a model trained based on pre-set labels associated with user preferences and used to predict the recommendation index of a video for a specific user; A sorting unit, used to sort the candidate videos in the candidate video set from large to small according to the recommendation index to obtain a recommendation list; A pushing unit is used to push the recommendation list to the target user.

13. The device according to claim 12, wherein: The tags include at least two play time tags, each of which indicates the duration of a user's preference for a video.

14. The device according to claim 13, wherein: The prediction unit comprises: A preprocessing module, used to preprocess the historical viewing data and each candidate video in the candidate video set to obtain target user information and video information corresponding to each candidate video; A first determination module is used to determine a preferred duration corresponding to the target user based on the target user information, wherein the preferred duration includes a user preferred playback duration corresponding to at least one video type; A calculation module is used to calculate, for each candidate video, the scores of multiple tags corresponding to the candidate video according to the video information corresponding to the candidate video, the target user information and the preferred duration, through the refined sorting multi-objective model to obtain a score corresponding to each tag; The second determination module is used to determine the recommendation index of the candidate video based on the scores corresponding to all tags.

15. The device according to claim 14, wherein: The computing module comprises: A determination submodule, used to determine, for each candidate video, the video type of the candidate video according to the video information corresponding to the candidate video; A first calculation submodule is used to calculate the label score of each playback duration corresponding to the candidate video through the refined sorting multi-objective model based on the video information corresponding to the candidate video, the target user information and the user preferred playback duration corresponding to the video type; The second calculation submodule is used to calculate other label scores through the refined ranking multi-objective model according to the video information corresponding to the candidate video and the target user information.

16. The device according to claim 14, wherein: The second determining module includes: A determination submodule, used to determine, for each tag, a weight corresponding to the tag; The calculation submodule is used to perform weighted fusion calculation on the scores corresponding to all tags based on the weight corresponding to each tag, so as to obtain the recommendation index of the candidate video.

17. The device according to claim 13, wherein: The tag also includes at least one of the following tags: Satisfaction label, completion rate label, resource debiasing label, mid-broadcast label, finished broadcast label, and fast-swipe label.

18. A model training device, comprising: A calibration unit, configured to calibrate historical viewing data corresponding to a plurality of users based on tags associated with user preferences to obtain a plurality of training samples, wherein the plurality of training samples include positive samples indicating recommendation and negative samples indicating recommendation; A division unit, used for dividing the plurality of training samples into a training set and a test set; The training unit is used to iteratively train the pre-built multi-objective model according to the training set, the test set and the label to obtain a refined multi-objective model for predicting the recommendation index of the video for a specific user.

19. The device according to claim 18, wherein: The tags include at least two play time tags, each of which indicates the duration of a user's preference for a video.

20. The apparatus of claim 18, further comprising: A construction unit, used to construct the multi-objective model, wherein the multi-objective model includes a shared underlying network architecture and a plurality of towers; Among them, each label corresponds to a tower.

21. The device according to claim 20, wherein: The training unit comprises: The training module is used to iteratively train the tower corresponding to each label based on the training set and the test set to obtain the refined multi-objective model.

22. The apparatus of claim 19, wherein the tag further comprises at least one of the following tags: Satisfaction label, completion rate label, resource debiasing label, mid-broadcast label, finished broadcast label, and fast-swipe label.

23. An electronic device, comprising: at least one processor; as well as a memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the method according to any one of claims 1 to 11.

24. A non-transitory computer-readable storage medium storing computer instructions, wherein: The computer instructions are used to cause the computer to execute the method according to any one of claims 1-11.

25. A computer program product, comprising a computer program, which, when executed by a processor, implements the steps of the method according to any one of claims 1 to 11.

Images