Video recommendation method and device

Abstract

The invention relates to the technical field of video recommendation, in particular to a video recommendation method and device, and the method comprises the steps: obtaining the identification information of a target user and a video user set, and the target user is a user in the video user set; based on the identification information of the target user and a pre-constructed long-term interest map, obtaining a trained long-term interest model of the target user, the long-term interest map being used for identifying the long-term interest preference of each user in the video user set; based on the identification information of the target user and a pre-constructed short-term interest map, obtaining a trained short-term interest model of the target user, the short-term interest map being used for identifying short-term interest preference of each user in the video user set; and performing video recommendation for the target user based on the long-term interest map, the short-term interest map, the long-term interest model and the short-term interest model. According to the method, more personalized video recommendation can be provided, personalized requirements of users can be met, and dependence on a large number of computing resources can be reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of video recommendation, and in particular, to a video recommendation method and apparatus. Background Art

[0002] The video recommendation system for intelligent large TV screens is one of the important directions in the development of current intelligent TV technology. Through complex algorithms and intelligent analysis, the system recommends video content that meets the interests and preferences of users, thereby providing a more personalized and intelligent viewing experience.

[0003] However, when the existing video recommendation system makes video recommendations, since a large amount of computing resources need to be mobilized for calculation, it not only increases the burden on the system operation but also makes it difficult to ensure the immediacy of recommendations. Summary of the Invention

[0004] Based on the above technical problems, the present invention provides a video recommendation method and apparatus.

[0005] According to one aspect of the present invention, there is provided a video recommendation method, including: Obtaining the identification information of a target user and a video user set, where the target user is a user in the video user set; Based on the identification information of the target user and a pre-constructed long-term interest graph, obtaining a trained long-term interest model of the target user, where the long-term interest graph is used to identify the long-term interest preferences of each user in the video user set; Based on the identification information of the target user and a pre-constructed short-term interest graph, obtaining a trained short-term interest model of the target user, where the short-term interest graph is used to identify the short-term interest preferences of each user in the video user set; Based on the long-term interest graph, short-term interest graph, long-term interest model, and short-term interest model, making video recommendations for the target user.

[0006] According to the video recommendation method of one aspect of the present invention, the construction method of the long-term interest graph includes: Obtaining a video user set, a video set, and historical behavior log data of the video user set; Based on the video user set, the video set, and the historical behavior log data of the video user set, determining a first interaction relationship between the video user set and the video set; Based on the video user set, the video set, and the first interaction relationship, constructing a long-term interest graph.

[0007] According to the video recommendation method of one aspect of the present invention, based on the identification information of the target user and a pre-constructed long-term interest graph, obtaining a trained long-term interest model of the target user includes: Constructing a long-term interest model; Based on the identification information of the target user, determine the first information of the target user and the first video information related to the first information from the pre-constructed long-term interest graph, where the first information is the historical portrait of the target user; Based on the first information and the first video information, train the long-term interest model to obtain the trained long-term interest model of the target user.

[0008] According to the video recommendation method of one aspect of the present invention, the construction method of the short-term interest graph includes: Obtain the video user set, the video set, and the recent behavior log data of the video user set; Based on the video user set, the video set, and the short-term behavior log data of the video user set, determine the second interaction relationship between the video user set and the video set; Based on the video user set, the video set, and based on the second interaction relationship, construct the short-term interest graph.

[0009] According to the video recommendation method of one aspect of the present invention, based on the identification information of the target user and the pre-constructed short-term interest graph, obtain the trained short-term interest model of the target user, including: Construct the short-term interest model; Based on the identification information of the target user, determine the second information of the target user and the second video information related to the second information from the pre-constructed short-term interest graph, where the second information is the recent portrait of the target user; Based on the second information and the second video information, train the short-term interest model to obtain the trained short-term interest model of the target user.

[0010] According to the video recommendation method of one aspect of the present invention, based on the long-term interest model and the short-term interest model, perform video recommendation for the target user, including: Based on the short-term interest graph, obtain the third information of the target user at the current moment and the third video information related to the third information of the target user, where the third information is the current portrait of the target user; Input the third information and the third video information into the short-term interest model to obtain the first video recommendation list output by the short-term interest model; Based on the long-term interest graph, obtain the fourth information of the target user at the historical moment and the fourth video information related to the fourth information of the target user, where the fourth information is the portrait of the target user within half a year to one year; Input the fourth information and the fourth video information into the long-term interest model to obtain the second video recommendation list output by the long-term interest model; Based on the first video recommendation list and the second video recommendation list, perform video recommendation for the target user.

[0011] According to the video recommendation method of one aspect of the present invention, the construction method of the video set includes: Obtain a set of video items in the video library that are at least related to the video user set; Preprocess the set of video items; Extract feature points for each video in the preprocessed set of video items, and obtain first feature points that meet the first preset condition in each video; Based on the first feature points, determine a set of first feature maps in each video that contain the first feature points; Perform network processing on the set of first feature maps to obtain a set of second feature maps; Compare the similarity of every two adjacent frames of second feature maps in the set of second feature maps, and obtain a set of third feature maps that meet the second preset condition; Based on the set of third feature maps of each video in the preprocessed set of video items, obtain a video set.

[0012] According to the video recommendation method of one aspect of the present invention, the first preset condition includes: The absolute difference between the gray value of the surrounding pixel points of the first feature point and the gray value of the central pixel point of the first feature point is less than a preset first threshold, and the number of surrounding pixel points of the first feature point exceeds the sampling threshold; The second preset condition includes: The difference between every two adjacent frames of second feature maps exceeds a preset second threshold.

[0013] According to the video recommendation method of one aspect of the present invention, the method further includes: Obtain the viewing behavior record of the target user at the current moment and the previous viewing behavior record; Compare the similarity of the video information in the viewing behavior record of the target user at the current moment and the video information in the previous viewing behavior record. If the degree of content change of the video information in the viewing behavior record at the current moment exceeds a preset third threshold, update the short-term interest map.

[0014] According to another aspect of the present invention, there is provided a video recommendation device, including: An identification information acquisition module, configured to acquire the identification information of the target user and the video user set, where the target user is a user in the video user set; A long-term interest model determination module, configured to obtain a trained long-term interest model of the target user based on the identification information of the target user and a pre-constructed long-term interest map, where the long-term interest map is used to identify the long-term interest preferences of each user in the video user set; A short-term interest model determination module, configured to obtain a trained short-term interest model of the target user based on the identification information of the target user and a pre-constructed short-term interest map, where the short-term interest map is used to identify the short-term interest preferences of each user in the video user set; A video recommendation module for performing video recommendations for a target user based on a long-term interest graph, a short-term interest graph, a long-term interest model, and a short-term interest model.

[0015] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored on the memory and running on the processor. When the processor executes the computer program, it implements any one of the above video recommendation methods.

[0016] The present invention also provides a non-transitory computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, it implements any one of the above video recommendation methods.

[0017] The present invention also provides a computer program product, including a computer program. When the computer program is executed by a processor, it implements any one of the above video recommendation methods.

[0018] The video recommendation method and device provided by the present invention determine the long-term and short-term interest models of a target user through the identification information, long-term interest graph, and short-term interest graph of the target user, and combine the long-term and short-term interest models of the target user with the long-term interest graph and short-term interest graph for video recommendation. It can not only provide more personalized video recommendations to meet the personalized needs of users, but also reduce the dependence on a large amount of computing resources, improve the efficiency of the recommendation process, and combine with the short-term interest model to quickly respond to the latest behavior of users and improve the timeliness of the recommendation results. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the following drawings are some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0020] Figure 1 is one of the flow diagrams of the video recommendation method provided by the present invention.

[0021] Figure 2 is the second flow diagram of the video recommendation method provided by the present invention.

[0022] Figure 3 is one of the diagrams of the intelligent TV large-screen video recommendation system applying the video recommendation method provided by the present invention.

[0023] Figure 4 is the second diagram of the intelligent TV large-screen video recommendation system applying the video recommendation method provided by the present invention.

[0024] Figure 5 It is the third schematic diagram of the intelligent TV large-screen video recommendation system applying the video recommendation method provided by the present invention.

[0025] Figure 6 It is the fourth schematic diagram of the intelligent TV large-screen video recommendation system applying the video recommendation method provided by the present invention.

[0026] Figure 7 It is the third schematic diagram of the process of the video recommendation method provided by the present invention.

[0027] Figure 8 It is the fifth schematic diagram of the intelligent TV large-screen video recommendation system applying the video recommendation method provided by the present invention.

[0028] Figure 9 It is the structural schematic diagram of the video recommendation device provided by the present invention. Detailed implementation manners

[0029] To make the objectives, technical solutions and advantages of the present invention clearer, the technical solutions in the present invention will be clearly and completely described below with reference to the accompanying drawings in the present invention. Obviously, the described embodiments are some but not all of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments in the present invention without creative efforts shall fall within the protection scope of the present invention.

[0030] The intelligent TV large-screen video recommendation system is one of the important directions in the development of current smart TV technology. This system recommends video content that meets the interests and preferences of users through complex algorithms and intelligent analysis, thereby providing a more personalized and intelligent viewing experience. When the intelligent TV large-screen is powered on or the user enters the home page, it will automatically generate a personalized video recommendation list according to the user's interest preferences and viewing history; this recommendation method allows users to see the content they are interested in as soon as they turn on the TV, improving the viewing efficiency.

[0031] Existing intelligent large-screen video recommendation applications have the following defects: First, the current research focus is too concentrated on improving the accuracy of video recommendations, while insufficient consideration is given to the immediacy of recommendations. This directly leads to a discount in the user experience during viewing, making it difficult for users to enjoy smooth and immediate content recommendation services. Second, when the existing system makes video recommendations, it faces the problem of excessive consumption of computing resources. Each recommendation calculation requires mobilizing a large amount of computing resources, which not only increases the burden on the system operation but may also affect the smooth operation of other functions, restricting the overall efficiency of the system. Third, when traditional methods process short video feature extraction, they adopt a relatively mechanical way, that is, regarding each frame image of the short video as an isolated unit for individual recognition. Although this approach is meticulous, in the entire short video recognition process, it inevitably causes a waste of a large amount of repeated calculations, significantly increasing the computational complexity and directly leading to a decline in processing efficiency. This inefficient processing method is obviously difficult to meet the requirements in the modern era of pursuing high-speed and efficient user experiences.

[0032] Based on the above technical problems, the present invention provides a video recommendation method.

[0033] Figure 1 It is one of the flow schematic diagrams of the video recommendation method provided by the present invention.

[0034] As Figure 1 shown, the video recommendation method provided by the embodiments of the present invention specifically includes the following steps: Step 101, obtain the identification information of the target user and the video user set, where the target user is a user in the video user set.

[0035] In an embodiment of the present invention, the target user refers to the user for whom video recommendations are to be made. The video user set refers to all user groups that interact with video content. Among them, the target user is a user in the video user set. The identification information of the target user refers to the information used to uniquely identify the target user, such as: user account, device ID, or other user identity identification data. When making video recommendations, it is necessary to first obtain the identification information of the target user and the video user set. This helps to make video recommendations for the target user in the video user set in a targeted manner.

[0036] Step 102, based on the identification information of the target user and the pre-constructed long-term interest graph, obtain the trained long-term interest model of the target user, where the long-term interest graph is used to identify the long-term interest preferences of each user in the video user set.

[0037] In one embodiment of the present invention, the long-term interest graph includes the long-term interest preferences of all user groups interacting with the video content, and is used to represent the interest changes of users. The data in the long-term interest graph is derived from the long-term behavior patterns of users, such as historical viewing records, search history, video ordering history, likes, comments, etc. The target user is a member of all user groups interacting with the video content. By using the identification information of the target user to retrieve the information related to the target user in the pre-constructed long-term interest graph, a long-term interest model is trained, and this long-term interest model can reflect the long-term interest preferences of the target user, thereby obtaining the trained long-term interest model of the target user.

[0038] Step 103: Based on the identification information of the target user and the pre-constructed short-term interest graph, obtain the trained short-term interest model of the target user. The short-term interest graph is used to identify the short-term interest preferences of each user in the video user set.

[0039] In one embodiment of the present invention, the short-term interest graph includes the short-term interest preferences of all user groups interacting with the video content, and is used to represent the interest changes of users. The data of the short-term interest graph is derived from the videos recently watched by users, the keywords recently searched, the recent ordering history, likes, comments, etc. The target user is a member of all user groups interacting with the video content. By using the identification information of the target user to retrieve the information related to the target user in the short-term interest graph, a short-term interest model is trained, and this short-term interest model can reflect the short-term interest preferences of the target user, thereby obtaining the trained short-term interest model.

[0040] Step 104: Based on the long-term interest graph, short-term interest graph, long-term interest model, and short-term interest model, perform video recommendations for the target user.

[0041] In one embodiment of the present invention, the long-term interest graph, short-term interest graph, long-term interest model, and short-term interest model are combined to perform video recommendations for the target user. It not only considers the long-term interest preferences of the target user but also the user's recent interest preferences, thereby being able to provide more accurate and timely video recommendations for the target user.

[0042] In summary, according to the technical solution provided by the embodiments of the present invention, the long-term and short-term interest models of the target user are determined through the identification information of the target user, the long-term interest graph, and the short-term interest graph. The long-term and short-term interest models of the target user, along with the long-term interest graph and short-term interest graph, are combined for video recommendations. This can not only provide more personalized video recommendations to meet the personalized needs of users, but also reduce the dependence on a large amount of computing resources, improve the efficiency of the recommendation process, and by combining the short-term interest model, it can quickly respond to the user's latest behavior and improve the timeliness of the recommendation results.

[0043] Furthermore, the method for constructing a long-term interest graph includes: Obtain a video user set, a video set, and historical behavior log data of the video user set; Based on the video user set, the video set, and the historical behavior log data of the video user set, determine a first interaction relationship between the video user set and the video set; Based on the video user set, the video set, and the first interaction relationship, construct a long-term interest graph.

[0044] In an embodiment of the present invention, the video user set refers to all user groups that interact with video content. The video set refers to a set of video content available for users to watch. The historical behavior log data records the interaction behaviors of the users in the video user set with video content in the past period, such as viewing, searching, ordering, liking, commenting, etc. The first interaction relationship refers to the historical interaction relationship between the users in the video user set and the video set. The first interaction relationship includes playing, liking, ordering, collecting, etc. When constructing a long-term interest graph, it is necessary to obtain the video user set, the video set, and the historical behavior log data of the video user set, and analyze and determine the first interaction relationship between the users in the video user set and the video content in the video set in combination with the video user set, the video set, and the historical behavior log data of the video user set. Using the determined first interaction relationship, construct a long-term interest graph. The long-term interest graph connects the users in the video user set and the video content in the video set through their interaction relationships, forming a network structure, so that the long-term interest preferences of the users in the video user set can be analyzed. Exemplarily, when constructing a long-term interest graph, time can be used as an important dimension, and the interaction relationships between the users in the video user set and the video content in the video set are connected in chronological order. In this way, the long-term interest graph can not only show the relationship between users and video content, but also show the evolution of this relationship over time.

[0045] In summary, according to the technical solution provided by the embodiment of the present invention, the long-term interest graph can provide a comprehensive view to show the long-term interaction pattern between the users in the video set and the video content in the video set. Such a graph is crucial for understanding and predicting the long-term interests of users, because it can help the recommendation system capture the stable interest preferences of users, so as to provide more accurate long-term video recommendations.

[0046] Furthermore, based on the identification information of the target user and the pre-constructed long-term interest graph, obtaining a trained long-term interest model of the target user includes: Construct a long-term interest model; Based on the identification information of the target user, determine the first information of the target user and the first video information related to the first information from a pre-constructed long-term interest graph, where the first information is the historical portrait of the target user; Based on the first information and the first video information, train the long-term interest model to obtain the trained long-term interest model of the target user.

[0047] In an embodiment of the present invention, the first information refers to the historical portrait of the target user, including the long-term behavior data of the user, such as the types, frequencies, likes, comments, etc. of the videos watched in the past period of time. These data can be extracted from the long-term interest graph. At the same time, it is also necessary to determine the video information related to these long-term behavior data (i.e., "the first video information"), such as the metadata of the video (such as category, label, publisher, etc.). Create a long-term interest model, use the identification information of the target user to retrieve in the long-term interest graph, obtain the first information and the first video information, and use the first information of the target user and the relevant first video information to train the long-term interest model to analyze the historical behavior of the target user and the characteristics of the video content, so as to learn the long-term interest preferences of the target user. After training, the long-term interest model can predict the new video content that the user may be interested in according to the user's historical behavior.

[0048] In summary, according to the technical solution provided by the embodiment of the present invention, by constructing and training a model that can reflect the long-term interest preferences of users, it can be used for personalized video recommendation to improve the accuracy of recommendation and user satisfaction.

[0049] Further, the construction method of the short-term interest graph includes: Obtain the video user set, the video set, and the recent behavior log data of the video user set; Based on the video user set, the video set, and the short-term behavior log data of the video user set, determine the second interaction relationship between the video user set and the video set; Based on the video user set, the video set, and based on the second interaction relationship, construct the short-term interest graph.

[0050] In an embodiment of the present invention, the video user set refers to all user groups that interact with video content. The video set refers to a collection of video content available for users to watch. The recent behavior log data records the interaction behaviors of the users in the video user set with video content within a recent period of time, such as viewing, searching, ordering, liking, commenting, etc. The second interaction relationship refers to the interaction behaviors between the video user set and the video set, including playing, liking, ordering, collecting, etc. When constructing a short-term interest graph, it is necessary to first obtain the video user set, the video set, and the short-term behavior log data of the video user set, and determine the second interaction relationship between the video user set and the video set in combination with the video user set, the video set, and the short-term behavior log data of the video user set. Then, use the second interaction relationship to connect users and video content through their recent interaction behaviors to form a short-term interest graph, so as to analyze the short-term interest preferences of users. Exemplarily, when constructing a short-term interest graph, time can be used as an important dimension to connect the interaction relationships between the users in the video user set and the video content in the video set in chronological order. In this way, the short-term interest graph can not only show the relationship between users and video content, but also show the evolution of this relationship over time.

[0051] In summary, according to the technical solution provided by the embodiment of the present invention, the short-term interest graph can provide a dynamic view to show the short-term interaction patterns between the users in the video set and the video content in the video set. Such a graph is crucial for quickly capturing and responding to the short-term interest changes of users, because it can help the recommendation system timely adjust the recommendation strategy to adapt to the rapid changes of user interests. It can provide timely user interest feedback for the video recommendation system, thereby improving the timeliness and accuracy of recommendations.

[0052] Further, based on the identification information of the target user and the short-term interest graph, obtain the trained short-term interest model of the target user, including: Construct a short-term interest model; Based on the identification information of the target user, determine the second information of the target user and the second video information related to the second information from the short-term interest graph, and the second information is the recent portrait of the target user; Based on the second information and the second video information, train the short-term interest model to obtain the trained short-term interest model of the target user.

[0053] In one embodiment of the present invention, the "second information" refers to the recent portrait of the target user, which includes the user's recent behavioral data, such as the videos recently watched, the keywords searched, the likes and comments, etc. These data can be extracted from the short-term interest graph, which records the recent interaction relationship between the user and the video content. At the same time, it is also necessary to determine the video information related to these recent behavioral data ("second video information"), such as the metadata of the video (such as category, tag, publisher, etc.). Create a short-term interest model, use the identification information of the target user to retrieve in the short-term interest graph, obtain the second information and the second video information, and use the first information of the target user and the relevant first video information to train the short-term interest model to analyze the recent behavior of the target user and the characteristics of the video content, so as to learn the short-term interest preferences of the target user. After training, the short-term interest model can predict the new video content that the user may be interested in according to the user's recent behavior.

[0054] Exemplarily, obtain the identification information of the target user, such as the user ID. According to the user identification information, extract the second information (the recent portrait of the user) of the user and the second video information (the videos recently watched by the user) related to the second information from the pre-constructed short-term interest graph. Construct an NSSR model, and initialize the parameters of the NSSR model (i.e., the short-term interest model), including the embedding vectors of the user and the video, the parameters of the self-attention mechanism, etc. Use the second information and the second video information of the target user as the input of the model, and train the NSSR model. During the training process, use the validation set to evaluate the performance of the model. Until the performance of the NSSR model on the validation set reaches a satisfactory level or reaches the preset number of iterations. Obtain the trained NSSR model.

[0055] In summary, according to the technical solution provided by the embodiment of the present invention, by constructing and training a model that can reflect the short-term interest preferences of users, it can be used for personalized video recommendation, improving the timeliness and user satisfaction of the recommendation.

[0056] Furthermore, based on the long-term interest graph, the short-term interest graph, the long-term interest model, and the short-term interest model, video recommendation is performed for the target user, including: Based on the short-term interest graph, obtain the third information of the target user at the current moment and the third video information related to the third information of the target user. The third information is the current portrait of the target user; Input the third information and the third video information into the short-term interest model to obtain the first video recommendation list output by the short-term interest model; Based on the long-term interest graph, obtain the fourth information of the target user at the historical moment and the fourth video information related to the fourth information of the target user. The fourth information is the portrait of the target user within half a year to one year; Based on the fourth information and the fourth video information being input into the long-term interest model, obtain the second video recommendation list output by the long-term interest model; based on the first video recommendation list and the second video recommendation list, perform video recommendations for the target user.

[0057] In an embodiment of the present invention, the "third information" refers to the portrait of the target user at the current moment, including the user's real-time behavior data, such as the currently watched video, recent search queries, recent likes and comments, etc. The "third video information" is associated with the third information and refers to the detailed information of the video content that the user is currently interacting with. Obtain the third information and the third video information from the short-term interest graph, and input the third information and the third video information into the already trained short-term interest model. The short-term interest model will generate a first video recommendation list based on this information, and this first video recommendation list reflects the video content that the user may be interested in in the short term. Exemplarily, the third information can be the real-time behavior data at the recent time node, and the number of such real-time behavior data can be determined according to business requirements. Generally, it can be 20 - 80 pieces of real-time behavior data at the recent time node. The "fourth information" refers to the user portrait of the target user in the past period (such as half a year to one year), which includes the user's behavior data during this period. The "fourth video information" is associated with the fourth information and refers to the detailed information of the video content that the user interacted with during this period. Obtain the fourth information and the fourth video information, and input the fourth information and the fourth video information into the already trained long-term interest model. The long-term interest model will generate a second video recommendation list based on this information, and this second video recommendation list reflects the video content that the user is interested in in the long term. Combine the first video recommendation list generated by the short-term interest model and the second video recommendation list generated by the long-term interest model to perform comprehensive video recommendations for the target user. Such recommendations consider the immediate interest and long-term interest of the target user, aiming to provide more comprehensive and accurate personalized recommendations.

[0058] Exemplarily, weight distribution is performed on the first video recommendation list and the second video recommendation list. Larger weights are assigned to the video content of short-term interest to ensure that the recommended content is consistent with the user's current interest points. While smaller weights are assigned to the video content of long-term interest to maintain the diversity and long-term relevance of the recommended content. Fuse the first video recommendation list and the second video recommendation list to perform comprehensive video recommendations for the target user, which can balance the short-term interest and long-term interest of the target user and provide more comprehensive recommendations.

[0059] Exemplarily, user information (i.e., information of the video set user), video set information, and interaction information between the video set user and the video set are extracted from the long-term interest graph and the short-term interest graph. These data will serve as the basis for training the neural network model. According to the collected data, features are designed to reflect the interests of users and the characteristics of videos. For example, it can include the user's viewing history, the popularity of the video, the user's rating of the video, etc. A neural network model, such as a deep learning neural network collaborative filtering model (NCF), is used to learn the association between the interests of users and the characteristics of videos. The model will learn from the training data how to predict the degree of preference of users for unviewed videos. Using the trained neural network model, the video content in the first video recommendation list and the second video recommendation list is scored. This score reflects the degree of match between the video and the user's current interests. The videos in the two recommendation lists are sorted according to the model scores, or the two lists are merged into a comprehensive recommendation list through a specific fusion algorithm (such as weighted average, rank fusion, etc.). The possibility of interaction between the user and the un-interacted videos is predicted through the videos with which the user has already interacted. This can be achieved through the scores output by the model, and the higher the score, the greater the predicted possibility of user interaction.

[0060] For example, assume that a user often watches science fiction movies. The long-term interest graph may contain all the science fiction movies the user has watched in the past year, while the short-term interest graph may contain the viewing records in the most recent month. By analyzing this data, the neural network model can learn the user's preference for science fiction movies and recommend new science fiction movies to the user in the first video recommendation list and the second video recommendation list. The model will score the videos in the recommendation list according to the degree of the user's preference for science fiction movies, and then make recommendations based on the scores.

[0061] Exemplarily, sorting can also be performed using a neural network model. Specifically, the feature maps (i.e., the following third feature maps) and feature points corresponding to the video data in the first video list and the second video list are respectively converted into numerical feature vectors that can be processed by the neural network. These numerical feature vectors will represent the video data and be used for subsequent sorting learning. The RankNet model is used for sorting learning (RankNet is a neural network model suitable for sorting tasks. It trains the model by comparing paired samples so that video data with higher relevance obtains a higher ranking). In the training phase, the model receives a pair of video feature vectors as input and outputs a predicted value of their relevance score. Then, the model uses a loss function to calculate the difference between the predicted value and the true relevance label and updates the model weights to minimize this difference. Once the model is trained and its performance is verified, it can be used to sort new video data. For each video, the model outputs a relevance score, and then the videos are sorted according to these scores. Videos with higher scores are considered more relevant and will therefore be ranked higher. According to the sorting results, the video data ranked higher can be recommended to the target user.

[0062] In summary, according to the technical solution provided by the embodiments of the present invention, by analyzing the current behavior and historical behavior of the target user, using short-term and long-term interest models to generate recommendation lists respectively, and combining these lists to provide video recommendations for the user, it is possible to balance the short-term interest changes and long-term stable interest preferences of the user, thereby improving the effect of the recommendation system and user satisfaction.

[0063] Figure 2 It is the second flowchart of the video recommendation method provided by the present invention.

[0064] As Figure 2 described, according to the video recommendation method provided by the embodiments of the present invention, wherein the method for constructing the video set specifically includes the following steps: Step 201, obtain a set of video items in the video library that are at least related to the video user set.

[0065] In an embodiment of the present invention, all video items at least related to the video user set are obtained from the video library to form a set of video items. Among them, the set of video items contains multiple videos.

[0066] Step 202, preprocess the set of video items.

[0067] In an embodiment of the present invention, the set of video items is preprocessed, and the preprocessing includes video format conversion, compression, and decoding preprocessing operations on the videos to facilitate subsequent feature extraction and analysis.

[0068] Step 203: Extract feature points from each video in the preprocessed video project set, and obtain first feature points that meet the first preset condition in each video.

[0069] In an embodiment of the present invention, feature points are extracted from each video in the preprocessed video project set to obtain first feature points that meet the first preset condition. The first feature points refer to key information points in the video. Exemplarily, feature points can be extracted by the corner detection method.

[0070] Further, the first preset condition includes: the absolute difference between the gray value of the surrounding pixel points of the first feature point and the gray value of the central pixel point of the first feature point is less than a preset first threshold, and the number of the surrounding pixel points of the first feature point exceeds the sampling threshold.

[0071] Exemplarily, feature points are extracted by the corner detection method and feature point determination is performed. The determination condition is: In the formula, G(p) represents the gray value of the central pixel point of the feature point, G(q) represents the gray value of the surrounding pixel points of the feature point, β is the preset first threshold, and F(G(p), G(q)) represents the comparison result of the gray values (i.e., the absolute difference). If the absolute difference between the gray value of the surrounding pixel points of the feature point and the gray value of the central pixel point of the feature point is less than the preset first threshold, and the number of the surrounding pixel points of the feature point exceeds the sampling threshold, then it is determined that the feature point is the first feature point that meets the first preset condition.

[0072] Step 204: Based on the first feature points, determine a first feature map set containing the first feature points in each video.

[0073] In an embodiment of the present invention, according to the feature point determination, a set of first feature points that meet the first preset condition in each video can be determined. Based on the determined first feature points, a first feature map set in each video is constructed. The first feature map set contains all the key feature points in the video, and these points can represent the key visual information of the video.

[0074] Step 205: Perform network processing on the first feature map set to obtain a second feature map set.

[0075] In an embodiment of the present invention, network processing is performed on the first feature map set. After grid processing, a second feature map set is formed.

[0076] Step 206: Compare the similarity between every two adjacent frames of the second feature map set in the second feature map set to obtain a third feature map set that meets the second preset condition.

[0077] In one embodiment of the present invention, the similarity between every two adjacent second feature maps in the second feature map set is compared to determine a third feature map set that meets the second preset condition.

[0078] Exemplarily, the second preset condition is that the difference between every two adjacent second feature maps exceeds a preset second threshold.

[0079] Specifically, the second preset condition refers to a preset second threshold set in the similarity comparison, which is used to determine whether the difference between two frames is large enough to be considered different. If the difference between two frames exceeds this preset second threshold, then they are considered dissimilar, which may indicate a scene change, the appearance of a new object, or other significant visual changes. Exemplarily, based on the second preset condition, the second feature maps that differ from the previous frame by more than the preset second threshold are selected to form the third feature map set. This third feature map set represents the key frames in the video, that is, the frames that best represent the changes in the video content.

[0080] Step 207: Based on the third feature map set of each video in the preprocessed video project set, obtain a video set.

[0081] In one embodiment of the present invention, the third feature map sets of each video in the preprocessed video project set constitute a video set.

[0082] In summary, according to the technical solution provided by the embodiments of the present invention, by steps such as preprocessing the video, extracting key feature points, constructing a feature map set, network processing, and similarity comparison, key frames are efficiently extracted from the video library, which can effectively reduce the dependence on computing resources during video recognition.

[0083] Furthermore, the method further includes: Obtain the viewing behavior record of the target user at the current moment and the previous viewing behavior record; Compare the video information of the viewing behavior record of the target user at the current moment with the video information of the previous viewing behavior record. If the degree of content change of the video information of the viewing behavior record at the current moment exceeds a preset third threshold, update the short-term interest map.

[0084] In one embodiment of the present invention, the viewing behavior records of the target user at the current moment and the previous viewing behavior records are obtained. These records may include video IDs viewed by the user, viewing durations, interaction behaviors, etc. The video information in the viewing behavior record of the user at the current moment is compared with the video information in the previous record for similarity. If the degree of content change in the video information of the viewing behavior record at the current moment exceeds a preset third threshold, the short-term interest map is updated. Exemplarily, the video information in the viewing behavior record of the user at the current moment (such as the feature points in the third feature map constituting the video information) is compared with the video information in the previous record (such as the feature points in the third feature map constituting the video information)). If the degree of content change in the video information of the viewing behavior record at the current moment compared with the previous one exceeds the preset third threshold, this indicates that the user's interest may have changed significantly. Among them, the third threshold is a preset standard for determining when the user's interest change is large enough to require updating the short-term interest map. Once it is determined that the user's interest has changed significantly, the short-term interest map is updated to reflect the user's latest interest preferences. This update helps the recommendation system adapt to the change of the user's interest faster and provide more accurate personalized recommendations.

[0085] Furthermore, by collecting user feedback, which can be obtained in various ways, such as the user's playback behavior (such as viewing duration, whether watched completely), ratings, likes, comments, or direct feedback options (such as the "dislike" button). The collected user feedback is used to evaluate the recommendation effect. Common evaluation metrics include click-through rate (CTR), conversion rate, user satisfaction, coverage rate, and diversity, etc. According to the evaluation results, the video recommendation list is optimized to improve the relevance and user satisfaction of the recommendation, ensuring that the recommended content not only meets the user's immediate needs but also can capture the user's long-term interests, thereby enhancing the overall user experience.

[0086] In summary, according to the technical solution provided by the embodiment of the present invention, by dynamically capturing the short-term changes of the user's interest and timely updating the user's interest model, the relevance and user satisfaction of the recommendation are improved. This method is particularly suitable for scenarios that need to quickly respond to changes in user behavior, such as real-time recommendation systems.

[0087] To illustrate in detail the video recommendation method provided by the embodiment of the present invention, refer to Figures 3 - 6 , Figure 3 which is one of the schematic diagrams of the intelligent TV large-screen video recommendation system applying the video recommendation method provided by the present invention. Figure 4 which is the second schematic diagram of the intelligent TV large-screen video recommendation system applying the video recommendation method provided by the present invention. Figure 5 which is the third schematic diagram of the intelligent TV large-screen video recommendation system applying the video recommendation method provided by the present invention.Figure 6 This is the fourth schematic diagram of the intelligent TV large-screen video recommendation system using the video recommendation method provided by the present invention.

[0088] like Figure 3 As shown, the intelligent large-screen TV video recommendation system includes a system management module 1, a data acquisition module 2, a video preprocessing model 3 and an intelligent recommendation module 4.

[0089] The system management module 1 includes a system monitoring module 11, a permission management module 12, a log recording module 13 and an anomaly detection module 14, and the system management module 1 establishes data connections with the video preprocessing module 3 and the intelligent recommendation module 4. The system management module 1, the system monitoring module 11, the permission management module 12, the log recording module 13 and the anomaly detection module 14 are used to manage the system and monitor and maintain the system stability. The system management module 1 establishes data connections with the data acquisition module 2, the video preprocessing module 3 and the intelligent recommendation module 4 respectively for system stability monitoring and maintenance. Specifically, the system management module 1 is used to control the data acquisition module 2 to collect data and determine the scope of data collection by the data acquisition module 2. The data acquisition module 2 includes a history record module 21, a search record module 22 and a behavior record module 23. The history record module 21, the search record module 22 and the behavior record module 23 are respectively used to collect the user's viewing history, search record, and behavior log data. The behavior recording module 23 includes the data collected by the history recording module 21 and the search recording module 22. In addition to collecting viewing history and search history, the behavior recording module 23 also collects other types of behavior data, such as ordering, clicking, sliding, page browsing, dwell time and other interactive behaviors. The data collection module 2 sends the data collected from the terminal used by the user of the video collection to the intelligent recommendation module 4. Figure 4 As shown, the video preprocessing module 3 includes a video decoding module 31 and a feature extraction module 32. The feature extraction module 32 includes an image feature module 321 and a verification and screening module 322, and the image feature module 321 and the verification and screening module 322 are used to extract image features and perform verification and screening. The video preprocessing module 3 receives data from the data acquisition module 2. The video preprocessing module 3 receives data from the data acquisition module 2 and sends the data to the intelligent recommendation module 4. The intelligent recommendation module 4 includes a feature matching module 41, a recall list module 42, a recommendation sorting module 43, an evaluation feedback module 44, a graph construction module 45, a user portrait module 46, a real-time recommendation module 47 and a cache optimization module 48. As shown Figure 5 As shown, the feature matching module 41 includes a similarity recognition module 411, a grid analysis module 412 and a change detection module 413. The feature matching module 41 is used to determine the degree of consistency between the video and the target feature.

[0090] likeFigure 6 As shown in Figure 6 , the graph construction module 45 includes a real-time recording module 451, a short-term interest module 452, a long-term interest module 453, and a sequence fusion module 454. The graph construction module 45 is used to construct a long-term interest graph and a short-term interest graph. The real-time recording module 451 processes each interaction behavior of the user with the video in real time and updates the user's short-term interest data. The long-term interest module 453 constructs a long-term interest model based on historical behavior log data and the long-term interest graph. The short-term interest module 452 constructs a short-term interest model based on short-term behavior log data and the short-term interest graph. The long-term interest model and the short-term interest model are updated in real time through the sequence fusion module 454.

[0091] To describe in detail the video recommendation method of the above intelligent TV large-screen video recommendation system, please refer to Figures 7 - 8 An embodiment provided by the present invention includes: Step 1, data collection; Step 2, feature extraction; Step 3, grid analysis and detection; Step 4, knowledge graph construction; Step 5, generation and sorting of the recommendation list; Step 6, real-time recommendation and update; Step 7, system optimization and feedback.

[0092] In the above Step 1, the historical record module 21, the search record module 22, and the behavior record module 23 in the data collection module 2 respectively collect the user's viewing history, search records, and behavior log data. The user's behavior log data is collected through the behavior record module 23, and the behavior log data includes the user's viewing history data, search records, etc. Or the user's viewing history data is collected through the historical record module 21 and the user's search records are collected through the search record module 22.

[0093] In the above Step 2, each video data in the video library is decoded and feature-extracted through the video preprocessing module 3. Specifically, the video decoding module 31 performs video format conversion, compression, and decoding preprocessing operations on the video data. After the operations are completed, the feature extraction module 32 extracts feature points from the video data. Exemplarily, feature points are extracted through the corner detection method, and feature point determination is performed. The determination condition is: In the formula, G(p) represents the gray value of the central pixel point of the feature point, G(q) represents the gray values of the pixel points around the feature point, β is a set threshold, and F(G(p), G(q)) represents the comparison result of the gray values. If the absolute difference between the gray value of the central pixel point of the feature point and the gray values of the pixel points around the feature point is less than the threshold (i.e., the aforementioned first threshold), and the number of pixel points around the feature point exceeds the sampling threshold, then the feature point is determined to be the first feature point (i.e., the first feature point that satisfies the first preset condition). Furthermore, each video data contains several first feature maps, and each feature map contains the first feature point.

[0094] In the above step 3, for each video data after extracting feature points, each frame of the video picture (i.e., the first feature map) of each video data is processed into a grid by the feature matching module 41, and the similarity between adjacent two frames is compared to match features. Specifically, the adjacent two-frame video pictures are processed into a grid by the grid analysis module 412, divided into L×L rectangular grids, and the proportion of unmatched feature points in each grid is calculated by the change detection module 413 to determine the degree of content change, and a determination threshold (i.e., the aforementioned second threshold) is set. The change degree is compared with the threshold to determine whether the video picture has a display change. The video picture with the change degree exceeding the threshold is taken as the key frame (i.e., the aforementioned third feature map) and then extracted. Among them, the third feature map corresponding to each video data and the feature points in the third feature map represent the video data, which can be used for subsequent map construction.

[0095] In the above step 4, the long-term interest map and the short-term interest map are constructed by the map construction module 45. The real-time recording module 451 processes each user interaction behavior in real time to update the user's short-term interest data in the short-term interest map. The long-term interest module 453 constructs a long-term interest model relying on the historical data of the user behavior log, and the short-term interest module 452 constructs a short-term interest model relying on the short-term data of the user behavior log. The long-term interest map of the user is maintained in the system without consuming time for each recommendation. If each time the user updates the interaction record, the user representation at the current moment (i.e., the aforementioned viewing behavior record) is compared with the user representation at the previous time (i.e., the aforementioned viewing behavior record) for similarity. Exemplarily, the video information corresponding to the user representation at the current moment is compared with the video information corresponding to the user representation at the previous time. If the similarity exceeds the threshold (i.e., the aforementioned third threshold), the short-term interest map is updated, and then the first video recommendation list output by the short-term interest model is updated. And the interest representation sequence (i.e., the aforementioned first video list and the second video list) is fused by the sequence fusion module 454 to update the recommendation list, which not only ensures the real-time nature of the recommendation but also avoids the problem of large computational complexity for each recommendation.

[0096] In the above step five, the long-term interest module 453 and the short-term interest module 452 generate and sort a recommendation list according to the first video list and the second video list through the recall list module 42 and the recommendation ranking module 43. Specifically, the recall list module 42 reads video data from the short-term interest graph and the long-term interest graph according to the first video list and the second video list, and performs deduplication and filtering. Then, the recommendation ranking module 43 sorts the video data in the deduplicated first video list and second video list, for example, sorts the videos based on time, to determine the final order recommended to the user, so as to finally recommend videos to the target user. Exemplarily, when sorting, it can be sorted in chronological order, that is, the ones with earlier time are set with a larger weight and sorted in the front of the recommendation.

[0097] Exemplarily, user information (i.e., information of the video set user), video set information, and interaction information between the video set user and the video set are extracted from the long-term interest graph and the short-term interest graph. These data will be used as the basis for training the neural network model. According to the collected data, features are designed to reflect the user's interests and the characteristics of the videos. For example, it can include the user's viewing history, the popularity of the videos, the user's ratings of the videos, etc. A neural network model, such as a deep learning neural network collaborative filtering model (NCF), is used to learn the association between the user's interests and the features of the videos. The model will learn how to predict the user's preference for unviewed videos through the training data. The trained neural network model is used to score the video content in the first video recommendation list and the second video recommendation list. This score reflects the degree of match between the video and the user's current interests. The videos in the two recommendation lists are sorted according to the model scores, or the two lists are merged into a comprehensive recommendation list through a specific fusion algorithm (such as weighted average, ranking fusion, etc.). The possibility of the user interacting with the un-interacted videos is predicted through the videos the user has already interacted with. This can be achieved through the scores output by the model, and the higher the score, the greater the predicted possibility of user interaction.

[0098] For example, assume that a user often watches science fiction movies. The long-term interest graph may contain all the science fiction movies the user has watched in the past year, while the short-term interest graph may contain the viewing records in the recent month. By analyzing these data, the neural network model can learn the user's preference for science fiction movies and recommend new science fiction movies to the user in the first video recommendation list and the second video recommendation list. The model will score the videos in the recommendation list according to the user's preference for science fiction movies, and then make recommendations according to the score ranking.

[0099] Exemplarily, it can also be sorting using a neural network model. Specifically, the feature maps (i.e., the aforementioned third feature maps) and feature points corresponding to the video data in the first video list and the second video list are respectively converted into numerical feature vectors that can be processed by the neural network. These numerical feature vectors will represent the video data and be used for subsequent sorting learning. The RankNet model is used for sorting learning (RankNet is a neural network model suitable for sorting tasks. It trains the model by comparing paired samples so that video data with higher relevance obtains a higher ranking). In the training phase, the model receives a pair of video feature vectors as input and outputs a predicted value of their relevance score. Then, the model uses a loss function to calculate the difference between the predicted value and the true relevance label, and updates the model weights to minimize this difference. Once the model is trained and its performance is verified, it can be used to sort new video data. For each video, the model outputs a relevance score, and then the videos are sorted according to these scores. Videos with higher scores are considered more relevant and will therefore be ranked higher. According to the sorting results, the video data ranked higher can be recommended to the target user.

[0100] In step six above, real-time video content recommendation is performed according to the long-term interest module 453 and the short-term interest module 452. Specifically, the long-term interest model constructed by the long-term interest module 453 and the short-term interest model constructed by the short-term interest module 452 combine the long-term interest graph and the short-term interest graph to recommend videos to the target user, realizing real-time video content recommendation.

[0101] In step seven above, the evaluation feedback module 44 collects user feedback, which can be obtained in various ways, such as the user's playback behavior (such as viewing duration, whether watched completely), rating, like, comment, or direct feedback options (such as the "dislike" button). The collected user feedback is used to evaluate the recommendation effect. Commonly used evaluation metrics include click-through rate (CTR), conversion rate, user satisfaction, coverage rate, and diversity, etc. According to the evaluation results, the work of the recall list module 42 and the recommendation sorting module 43 is optimized to improve the relevance and user satisfaction of the recommendation, ensuring that the recommended content meets both the user's immediate needs and can capture the user's long-term interests, thereby enhancing the overall user experience.

[0102] Based on the above, the advantages of the present invention are as follows: The data acquisition module 2 collects the user's viewing history, search records, and behavior log data, providing rich user behavior data for subsequent recommendations.

[0103] The video preprocessing module 3 decodes the video and extracts features. It extracts video feature points through methods such as corner detection, laying a foundation for video content analysis. The feature matching module 41 performs grid processing on the video frames and calculates the change degree between adjacent frames to identify key frames, providing support for the dynamic understanding of video content. This not only improves the video processing efficiency but also reduces the waste of computing resources. The graph construction module 45 constructs and updates the long-term interest graph and short-term interest graph in real time, capturing the long-term and short-term interest changes of users. The recall list module 42 and the recommendation ranking module 43 generate and rank the recommendation list according to the user's interest graph, providing personalized video recommendations. Furthermore, it realizes real-time video recommendations according to the long-term and short-term interest models of users, meeting the immediate viewing needs of users. The evaluation feedback module 44 collects user feedback, evaluates the recommendation effect, and optimizes the recommendation according to the feedback, improving the relevance of the recommendation and user satisfaction.

[0104] In summary, the video recommendation system realizes efficient, accurate, and personalized video recommendations by comprehensively utilizing user behavior data, video content features, and real-time feedback, aiming to improve the user experience and the user stickiness of the platform.

[0105] Figure 9 It is a schematic structural diagram of the video recommendation device provided by the present invention.

[0106] As Figure 9 shown, according to the video recommendation device 900 provided by the embodiment of the present invention, it includes: an identification information acquisition module 901, a long-term interest model determination module 902, a short-term interest model determination module 903, and a video recommendation module 904.

[0107] Furthermore, the identification information acquisition module 901 is used to acquire the identification information of the target user; Furthermore, the long-term interest model determination module 902 is used to obtain the trained long-term interest model of the target user based on the identification information of the target user and the long-term interest graph, and the long-term interest graph is used to identify the long-term interest preferences of the target user; Furthermore, the short-term interest model determination module 903 is used to obtain the trained short-term interest model of the target user based on the identification information of the target user and the short-term interest graph, and the short-term interest graph is used to identify the short-term interest preferences of the target user; Furthermore, the video recommendation module 904 is used to perform video recommendations for the target user based on the long-term interest graph, short-term interest graph, long-term interest model, and short-term interest model.

[0108] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored on the memory and running on the processor. When the processor executes the computer program, it implements any one of the above video recommendation methods.

[0109] The present invention also provides a non-transitory computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the above-mentioned video recommendation method is implemented.

[0110] The present invention also provides a computer program product, including a computer program, and when the computer program is executed by a processor, the above-mentioned video recommendation method is implemented.

[0111] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than limiting them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements on some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A video recommendation method, characterized in that: include: Obtain identification information of a target user and a video user set, wherein the target user is a user in the video user set; Based on the identification information of the target user and the pre-built long-term interest graph, a trained long-term interest model of the target user is obtained, wherein the long-term interest graph is used to identify the long-term interest preference of each user in the video user set; Based on the identification information of the target user and a pre-constructed short-term interest graph, a trained short-term interest model of the target user is obtained, wherein the short-term interest graph is used to identify the short-term interest preference of each user in the video user set; Based on the long-term interest graph, the short-term interest graph, the long-term interest model and the short-term interest model, video recommendations are made for the target user.

2. The video recommendation method according to claim 1, characterized in that: The method for constructing the long-term interest graph includes: Obtaining a video user set, a video set, and historical behavior log data of the video user set; Determining a first interaction relationship between the video user set and the video set based on the video user set, the video set, and historical behavior log data of the video user set; The long-term interest graph is constructed based on the video user set, the video set, and the first interactive relationship.

3. The video recommendation method according to claim 1, characterized in that: The step of acquiring a trained long-term interest model of the target user based on the identification information of the target user and a pre-built long-term interest graph includes: Build long-term interest models; Based on the identification information of the target user, determining first information of the target user and first video information related to the first information from the pre-constructed long-term interest graph, where the first information is a historical portrait of the target user; Based on the first information and the first video information, the long-term interest model is trained to obtain a trained long-term interest model of the target user.

4. The video recommendation method according to claim 1, characterized in that: The method for constructing the short-term interest graph includes: Obtaining a video user set, a video set, and recent behavior log data of the video user set; Determining a second interaction relationship between the video user set and the video set based on the video user set, the video set, and the short-term behavior log data of the video user set; The short-term interest graph is constructed based on the video user set, the video set and the second interactive relationship.

5. The video recommendation method according to claim 4, characterized in that: The step of acquiring a trained short-term interest model of the target user based on the identification information of the target user and a pre-built short-term interest graph includes: Construct short-term interest models; Based on the identification information of the target user, determining second information of the target user and second video information related to the second information from the pre-constructed short-term interest graph, where the second information is a recent portrait of the target user; Based on the second information and the second video information, the short-term interest model is trained to obtain a trained short-term interest model of the target user.

6. The video recommendation method according to claim 1, characterized in that: The recommending videos to the target user based on the long-term interest graph, the short-term interest graph, the long-term interest model and the short-term interest model includes: Based on the short-term interest graph, obtaining third information of the target user at a current moment and third video information related to the third information of the target user, wherein the third information is a current portrait of the target user; Inputting the third information and the third video information into the short-term interest model, and obtaining a first video recommendation list output by the short-term interest model; Based on the long-term interest graph, fourth information of the target user at a historical moment and fourth video information related to the fourth information of the target user are obtained, wherein the fourth information is a portrait of the target user within six months to one year; Based on the fourth information and the fourth video information input into the long-term interest model, a second video recommendation list output by the long-term interest model is obtained; based on the first video recommendation list and the second video recommendation list, video recommendations are made for the target user.

7. The video recommendation method according to claim 2, characterized in that: The method for constructing the video set includes: Acquire a set of video items in a video library that is at least related to the set of video users; Preprocessing the video item set; Extracting feature points from each video in the preprocessed video item set to obtain a first feature point in each video that meets a first preset condition; Based on the first feature points, determining a first feature graph set containing the first feature points in each video; Performing network processing on the first feature graph set to obtain a second feature graph set; Comparing the similarity of the second feature maps of every two adjacent frames in the second feature map set to obtain a third feature map set that meets the second preset condition; The video set is obtained based on the third feature map set of each video in the preprocessed video item set.

8. The video recommendation method according to claim 7, characterized in that: The first preset condition includes: An absolute difference between the grayscale value of the surrounding pixels of the first feature point and the grayscale value of the central pixel of the first feature point is less than a preset first threshold, and the number of surrounding pixels of the first feature point exceeds a sampling threshold; The second preset condition includes: The difference between the second feature maps of every two adjacent frames exceeds a preset second threshold.

9. The video recommendation method according to claim 7, characterized in that: The method further comprises: Obtaining the current viewing behavior record and the last viewing behavior record of the target user; The video information of the target user's current viewing behavior record and the video information of the last viewing behavior record are compared for similarity. If the content change degree of the video information of the current viewing behavior record exceeds a preset third threshold, the short-term interest map is updated.

10. A video recommendation device, characterized in that: include: An identification information acquisition module, used to acquire identification information of a target user and a video user set, wherein the target user is a user in the video user set; A long-term interest model determination module, used to obtain a trained long-term interest model of the target user based on the identification information of the target user and a pre-constructed long-term interest graph, wherein the long-term interest graph is used to identify the long-term interest preference of the target user; A short-term interest model determination module, used to obtain a trained short-term interest model of the target user based on the identification information of the target user and a pre-constructed short-term interest graph, wherein the short-term interest graph is used to identify the short-term interest preference of the target user; A video recommendation module is used to recommend videos to the target user based on the long-term interest graph, the short-term interest graph, the long-term interest model and the short-term interest model.

Images