A big data-based video recommendation method and system

Abstract

The application provides a video recommendation method and system based on big data, and relates to the technical field of video recommendation. The method comprises the following steps: inputting user chat records into a video output model to determine multiple discussion videos in the user chat records and discussion chat records corresponding to each discussion video; determining multiple close discussion friends and the closeness of the multiple close discussion friends to the user based on the multiple discussion videos in the user chat records and the discussion chat records corresponding to each discussion video by using a friend output model; constructing a graph structure; processing the graph structure based on a graph neural network model to determine multiple recommended videos; and recommending the user based on the multiple recommended videos. The method can improve the accuracy of video recommendation by using user social data, thereby improving user experience.

Description

Technical Field

[0001] This invention relates to the field of video recommendation technology, and specifically to a video recommendation method and system based on big data. Background Technology

[0002] With the rapid development of internet technology, watching videos has become an important part of people's daily lives. Big data-based video recommendation systems, acting as a bridge connecting users and video content, directly impact user viewing experience and platform user stickiness. Traditional big data-based video recommendation methods primarily rely on users' historical behavioral data, such as viewing records and search history, using algorithms like collaborative filtering, content recommendation, or hybrid recommendation to predict videos users might be interested in. However, these methods have some limitations. First, they often ignore information from users' social networks, such as interactions and discussions between users and their friends, which contain rich information about user interests and preferences. Video discussions on social networks often involve multifaceted information, such as emotions, opinions, and context; these complex social signals are difficult to fully utilize in traditional recommendation systems.

[0003] Therefore, how to leverage user social data to improve the accuracy of video recommendations and thus enhance user experience is a problem that needs to be solved. Summary of the Invention

[0004] The main technical problem this invention addresses is how to utilize user social data to improve the accuracy of video recommendations, thereby enhancing the user experience.

[0005] According to a first aspect, the present invention provides a video recommendation method based on big data, comprising: acquiring user chat history; inputting the user chat history into a video output model to determine multiple discussion videos in the user chat history and discussion chat history corresponding to each discussion video; using a friend output model to determine multiple close discussion friends and the closeness between the multiple close discussion friends and the user based on the multiple discussion videos in the user chat history and the discussion chat history corresponding to each discussion video; acquiring the video browsing history of each close discussion friend and the user's video browsing history; constructing a graph structure, the graph structure including multiple nodes and multiple edges between the multiple nodes, the multiple nodes including a user node and multiple close discussion friend nodes, the user node being the central node, the multiple close discussion friend nodes establishing multiple edges with the user node respectively, the node feature of the user node being the user's video browsing history, the node feature of each close discussion friend node being the video browsing history of each close discussion friend, and the edge feature of the multiple edges being the closeness between the multiple close discussion friends and the user; processing the graph structure based on a graph neural network model to determine multiple recommended videos; and recommending videos to the user based on the multiple recommended videos.

[0006] Furthermore, the step of recommending videos to users based on the multiple recommended videos includes: performing K-means clustering on the multiple recommended videos to obtain K clusters and a cluster center for each of the K clusters; obtaining the recommended video closest to the cluster center of each cluster to obtain the recommended video corresponding to each cluster; scoring the recommended video corresponding to each cluster based on a scoring model to obtain a rating for the recommended video corresponding to each cluster; arranging the ratings of the recommended videos corresponding to each cluster in descending order to obtain the display order of the recommended videos corresponding to each cluster; and recommending the recommended videos to users for playback according to the display order of the recommended videos corresponding to each cluster.

[0007] Furthermore, the step of recommending videos to users based on the multiple recommended videos includes: performing K-means clustering on the multiple recommended videos to obtain K clusters and a cluster center for each of the K clusters; obtaining the recommended video closest to the cluster center of each cluster to obtain the recommended video corresponding to each cluster; determining the closeness of the close discussion friends corresponding to the recommended videos corresponding to each cluster; arranging the closeness of the close discussion friends corresponding to the recommended videos corresponding to each cluster in descending order to obtain the display order of the recommended videos corresponding to each cluster; and recommending the recommended videos to users for playback according to the display order of the recommended videos corresponding to each cluster.

[0008] Furthermore, the multiple recommended videos are selected from the video browsing history of each of the close friends with whom they have discussed the topic.

[0009] Furthermore, the video output model is a Transformer model, the input of the video output model is the user chat history, and the output of the video output model is multiple discussion videos in the user chat history and the discussion chat history corresponding to each discussion video.

[0010] According to a second aspect, the present invention provides a video recommendation system based on big data, comprising:

[0011] The first acquisition module is used to acquire user chat history;

[0012] The video output module is used to input the user chat history into the video output model to determine multiple discussion videos in the user chat history and the discussion chat history corresponding to each discussion video;

[0013] The friend output module is used to determine multiple close discussion friends and the degree of closeness between the multiple close discussion friends and the user based on multiple discussion videos in the user's chat history and the discussion chat history corresponding to each discussion video using the friend output model.

[0014] The second acquisition module is used to acquire the video browsing history of each close friend and the user's video browsing history;

[0015] A construction module is used to construct a graph structure, which includes multiple nodes and multiple edges between the nodes. The multiple nodes include a user node and multiple close discussion friend nodes. The user node is the central node. The multiple close discussion friend nodes each establish multiple edges with the user node. The node feature of the user node is the user's video browsing history. The node feature of each close discussion friend node is the video browsing history of each close discussion friend. The feature of the multiple edges is the degree of closeness between the multiple close discussion friends and the user.

[0016] The recommended video determination module is used to process the graph structure based on a graph neural network model to determine multiple recommended videos;

[0017] The recommendation module is used to make recommendations to users based on the multiple recommended videos.

[0018] Furthermore, the recommendation module is also used to: perform K-means clustering on the multiple recommended videos to obtain K clusters and the cluster center of each of the K clusters; obtain the recommended video closest to the cluster center of each cluster to obtain the recommended video corresponding to each cluster; score the recommended video corresponding to each cluster based on a scoring model to obtain the rating of the recommended video corresponding to each cluster; arrange the ratings of the recommended videos corresponding to each cluster in descending order to obtain the display order of the recommended videos corresponding to each cluster; and recommend the recommended videos to the user for playback according to the display order of the recommended videos corresponding to each cluster.

[0019] Furthermore, the recommendation module is also used to: perform K-means clustering on the multiple recommended videos to obtain K clusters and the cluster center of each of the K clusters; obtain the recommended video closest to the cluster center of each cluster to obtain the recommended video corresponding to each cluster; determine the closeness of the close discussion friends corresponding to the recommended video corresponding to each cluster; arrange the closeness of the close discussion friends corresponding to the recommended video corresponding to each cluster in descending order to obtain the display order of the recommended videos corresponding to each cluster; and recommend the recommended videos to the user for playback according to the display order of the recommended videos corresponding to each cluster.

[0020] Furthermore, the multiple recommended videos are selected from the video browsing history of each of the close friends with whom they have discussed the topic.

[0021] Furthermore, the video output model is a Transformer model, the input of the video output model is the user chat history, and the output of the video output model is multiple discussion videos in the user chat history and the discussion chat history corresponding to each discussion video.

[0022] This invention provides a video recommendation method and system based on big data. The method includes: acquiring user chat history; inputting the user chat history into a video output model to determine multiple discussion videos and corresponding discussion chat records in the user chat history; using a friend output model to determine multiple close discussion friends and the closeness between the multiple close discussion friends and the user based on the multiple discussion videos and the corresponding discussion chat records in the user chat history; acquiring the video browsing history of each close discussion friend and the user's video browsing history; and constructing a graph structure, the graph structure including multiple nodes and multiple edges between the multiple nodes, the multiple nodes including... A user node and multiple close discussion friend nodes are used, with the user node as the central node. Each close discussion friend node has multiple edges connected to the user node. The node feature of the user node is the user's video browsing history, and the node feature of each close discussion friend node is the video browsing history of that close friend. The edge feature represents the degree of closeness between the close discussion friend and the user. A graph neural network model is used to process the graph structure to determine multiple recommended videos. Based on these recommended videos, recommendations are made to the user. This method utilizes user social data to improve the accuracy of video recommendations, thereby enhancing the user experience. Attached Figure Description

[0023] Figure 1 A flowchart illustrating a video recommendation method based on big data provided in an embodiment of the present invention;

[0024] Figure 2 This is a schematic diagram illustrating a process for recommending videos to users based on multiple recommended videos, provided as an embodiment of the present invention.

[0025] Figure 3 This is a schematic diagram illustrating another process for recommending content to users based on multiple recommended videos, provided as an embodiment of the present invention.

[0026] Figure 4 This is a schematic diagram of a video recommendation system based on big data, provided as an embodiment of the present invention. Detailed Implementation

[0027] In this embodiment of the invention, the following are provided: Figure 1 The above describes a video recommendation method based on big data, which includes steps S1 to S7:

[0028] Step S1: Obtain user chat history.

[0029] User chat history refers to the content of conversations generated when users communicate with others in a chat application. User chat history includes text, video, images, etc.

[0030] Step S2: Input the user chat history into the video output model to determine multiple discussion videos in the user chat history and the discussion chat history corresponding to each discussion video.

[0031] The video output model is a Transformer model. The input of the video output model is the user's chat history, and the output of the video output model is multiple discussion videos in the user's chat history and the discussion chat history corresponding to each discussion video.

[0032] The Transformer model consists of an encoder and a decoder. The encoder learns representations from the input sequence, incorporating a self-attention mechanism and a feed-forward network. The decoder, building upon the encoder, introduces an additional multi-head attention mechanism to decode the encoder output and generate the target sequence. The Transformer model can process video from user chat logs, better capturing the relationships within the temporal sequence of the video.

[0033] Multiple discussion videos in a user's chat history refer to video content from discussions between the user and their friends. For example, if a user sends a landscape video to a friend and then discusses it via text, this landscape video would be considered a discussion video.

[0034] Each discussion video's corresponding chat log refers to the text content in a user's chat history associated with that specific video. For example, for a discussion video, its corresponding chat log includes relevant conversations within the chat log related to the video's topic. For instance, if a user sends a landscape video to a friend and engages in a text discussion, the chat log for that landscape video might be "This lake is so beautiful!"

[0035] The Transformer model, through its self-attention mechanism, can simultaneously focus on information from different positions within the input sequence and performs exceptionally well in handling long-distance dependencies. This enables the model to identify video-related keywords or features in user chat logs, thereby determining the multiple discussion videos involved. The Transformer model combines visual and linguistic information, enabling visual-linguistic interaction to identify video content or related descriptions in user chat logs, thus identifying multiple discussion videos. The Transformer model can model and understand the context of user chat logs, thereby better grasping the content of user discussions. By capturing the associations between each word and its surrounding words, the model can more accurately determine the context of the discussion videos, thus identifying multiple discussion videos in the user chat log. Through its multi-head attention mechanism, the Transformer model can simultaneously focus on information from different positions and content. This allows the model to accurately identify the discussion chat log corresponding to each discussion video in the user chat log, thus distinguishing different video discussion content.

[0036] In some embodiments, the video output model includes a chat log initial screening layer, a discussion video output layer, and a chat log determination layer. All three layers include a Transformer structure. The input to the chat log initial screening layer is the user's chat log, and its output is a preliminary selection of video-related discussion chat logs. The input to the discussion video output layer is the preliminary selection of video-related discussion chat logs, and its output is multiple discussion videos. The input to the chat log determination layer is multiple discussion videos and the preliminary selection of video-related discussion chat logs, and its output is the discussion chat log corresponding to each discussion video. The chat log initial screening layer can initially filter video-related discussion chat logs from the user's chat log. The discussion video output layer can process the preliminary selection of video-related discussion chat logs and output multiple discussion videos from the user's chat log. The chat log determination layer can explicitly determine the discussion chat log corresponding to each discussion video, establishing a correspondence between videos and chat logs. By dividing the video output model into a chat log initial screening layer, a discussion video output layer, and a chat log confirmation layer, it is possible to quickly filter relevant chat logs, extract users' existing video resources, and establish a correspondence between videos and chat logs, thereby improving processing efficiency and accuracy.

[0037] Step S3: Based on the multiple discussion videos in the user's chat history and the discussion chat history corresponding to each discussion video, use the friend output model to determine multiple close discussion friends and the degree of closeness between the multiple close discussion friends and the user.

[0038] The friend output model is a Transformer model. The input of the friend output model is multiple discussion videos in the user's chat history and the discussion chat history corresponding to each discussion video. The output of the friend output model is multiple close discussion friends and the degree of closeness between the multiple close discussion friends and the user.

[0039] Multiple close discussion friends refer to the group of friends with whom a user most frequently interacts and discusses video content, identified through analysis and inference using the Transformer model based on the user's chat history and discussion video content.

[0040] The closeness of close discussion friends refers to the level of intimacy between users and close discussion friends in their video content interactions. For example, if a user frequently discusses video content with certain friends in their chat history, and these discussions and chat logs are input into a Transformer model, the model can analyze and identify the group of friends who interact with the user most frequently in video discussions. These friends can be identified as close discussion friends. The Transformer model will also determine the degree of closeness between multiple close discussion friends and the user based on the user's chat history and discussion video content. A higher degree of closeness indicates more frequent discussions between close discussion friends and the user regarding video content.

[0041] By analyzing discussion videos and corresponding chat logs in user chat history, the Transformer model can capture rich contextual information, including dialogue content, sentiment, and common topics, thereby inferring the friends with whom the user has close discussions and outputting the degree of closeness between the users and these close friends.

[0042] Step S4: Obtain the video browsing history of each close friend and the user's video browsing history.

[0043] The video browsing history of close friends is the record generated when close friends watch videos on the platform. The video browsing history of close friends includes information such as the content of the video watched, the duration of the watch, and the time of the watch.

[0044] A user's video browsing history is a record generated when the user watches videos on the platform. The user's video browsing history includes information such as the video content watched, the viewing duration, and the viewing time.

[0045] Step S5: Construct a graph structure, which includes multiple nodes and multiple edges between the nodes. The multiple nodes include a user node and multiple close discussion friend nodes. The user node is the central node. The multiple close discussion friend nodes each establish multiple edges with the user node. The node feature of the user node is the user's video browsing history. The node feature of each close discussion friend node is the video browsing history of each close discussion friend. The feature of the multiple edges is the degree of closeness between the multiple close discussion friends and the user.

[0046] A graph structure is a data structure composed of nodes and edges. Nodes include node features, which are attributes or characteristics that describe the properties or information of a node. Edges describe the relationships between two nodes. The feature of an edge is the degree of closeness between a user and their friends / friends.

[0047] Step S6: Process the graph structure based on the graph neural network model to determine multiple recommended videos.

[0048] A graph neural network (GNN) model consists of a graph neural network (GNN) and fully connected layers. A GNN is a type of neural network that operates directly on graph-structured data. The input to a GNN model is the graph structure, and the output is multiple recommended videos.

[0049] The recommended videos are those output by the graph neural network model and are used to recommend videos to users.

[0050] By constructing a graph structure, the relationship network between users and their close friends can be clearly reflected. This relationship information is crucial for personalized recommendations because users' interests and preferences are often influenced by their close friends. Using video browsing records of user nodes and close friend nodes as node features, and the degree of closeness as edge features, allows for more comprehensive utilization of data information. This helps the model better understand the interactions and influences between users and their friends, improving recommendation accuracy. Processing graph-structured data using graph neural network models can effectively learn the complex relationships and information transmission between nodes, thereby more accurately uncovering users' interests and social relationships. Compared to traditional recommendation algorithms, graph neural networks have better representation and learning capabilities when processing graph data.

[0051] Step S7: Recommend videos to users based on the multiple recommended videos.

[0052] In some embodiments, a scoring model can be used to filter the multiple recommended videos before making recommendations to users. Figure 2This invention provides a schematic diagram of a process for recommending users based on multiple recommended videos, comprising steps S21-S25:

[0053] Step S21: Perform K-means clustering on the multiple recommended videos to obtain K clusters and the cluster center of each of the K clusters.

[0054] In some embodiments, features of the multiple recommended videos can be extracted using a feature extraction model to obtain color histograms, optical flow, and Mel-frequency cepstral coefficients of the multiple recommended videos. Then, based on the color histograms, optical flow, Mel-frequency cepstral coefficients, and a pre-defined K value, a K-means clustering algorithm is used to obtain K clusters and a cluster center for each of the K clusters. Each cluster represents a class of similar recommended videos. Specifically, as an example: K recommended videos are randomly selected as initial cluster centers. For each recommended video, the distance between each recommended video and all initial cluster centers is calculated, and each recommended video is assigned to the nearest cluster center, forming K clusters. For each cluster, the average value of the feature vectors of all videos within the cluster is calculated, and this average value is used as the new cluster center. The above steps are repeated until the number of cluster centers reaches a predetermined number of iterations, at which point the clustering ends.

[0055] The feature extraction model is a Transformer model. The input of the feature extraction model is multiple recommended videos, and the output of the feature extraction model is the color histogram of multiple recommended videos, the optical flow of multiple recommended videos, and the Mel frequency cepstral coefficients of multiple recommended videos.

[0056] A color histogram for a video is used to describe the statistical characteristics of color distribution in the video. The color histogram divides the colors in the video into several intervals according to certain rules, then counts the number of pixels in each color interval, ultimately forming a histogram.

[0057] Optical flow in video describes the motion of pixels in an image over time. By calculating optical flow, the trajectory and velocity information of objects in the video can be obtained.

[0058] The Mel frequency cepstral coefficients of a video are coefficients used to represent the spectral characteristics of an audio signal.

[0059] In some embodiments, the value of K can be determined using a preset relationship table between the value of K and the user's average daily video viewing time. The longer the user's average daily video viewing time, the larger the value of K. The preset relationship table between the value of K and the user's average daily video viewing time is artificially constructed in advance.

[0060] Step S22: Obtain the recommended video that is closest to the cluster center of each cluster to get the recommended video corresponding to each cluster.

[0061] For each cluster, calculate the distance between the feature vector of each recommended video and the cluster center of that cluster. For each cluster, find the recommended video that is closest to the cluster center; this video will then be the recommended video for that cluster.

[0062] Step S23: Based on the scoring model, score the recommended videos corresponding to each cluster to obtain the score of the recommended videos corresponding to each cluster.

[0063] The scoring model is a Transformer model. The input to the scoring model is multiple recommended videos, and the output is the color histogram, optical flow, and Mel-frequency cepstral coefficients of the recommended videos. The scoring model evaluates the quality of the recommended videos corresponding to each cluster. By processing the recommended videos for each cluster, the scoring model assigns higher scores to videos with more refined visuals, higher clarity, and more attractive content. Higher scores for recommended videos within each cluster indicate higher video quality.

[0064] Step S24: Arrange the ratings of the recommended videos corresponding to each cluster in descending order to obtain the display order of the recommended videos corresponding to each cluster.

[0065] The recommended videos for each cluster are sorted from highest to lowest score to determine the display order of the recommended videos for each cluster.

[0066] Step S25: Recommend the recommended videos to the user for playback according to the display order corresponding to each cluster.

[0067] Once the display order of the recommended videos for each cluster is obtained, they are recommended to the user for playback.

[0068] In some embodiments, the multiple recommended videos can be filtered based on the closeness of the friends with whom the recommended videos are discussed before being recommended to the user. Figure 3 This is another schematic diagram of a process for recommending users based on multiple recommended videos, provided by an embodiment of the present invention. The process for recommending users based on multiple recommended videos includes steps S31 to S35:

[0069] Step S31: Perform K-means clustering on the multiple recommended videos to obtain K clusters and the cluster center of each of the K clusters.

[0070] For details on step S31, please refer to step S21, which will not be repeated here.

[0071] Step S32: Obtain the recommended video that is closest to the cluster center of each cluster to get the recommended video corresponding to each cluster.

[0072] For details on step S32, please refer to step S22; it will not be repeated here.

[0073] Step S33: Determine the closeness of the close discussion friends corresponding to the recommended video for each cluster.

[0074] Once the recommended videos for each cluster are determined, the closeness of the close discussion friends for each cluster's recommended videos is obtained based on the multiple close discussion friends identified in step S3 and the closeness between these close discussion friends and the user. The greater the closeness of the close discussion friends for each cluster's recommended videos, the closer the relationship between the user and that close discussion friend's video discussion, and thus that video can be prioritized for recommendation.

[0075] Step S34: Arrange the closeness of the close discussion friends corresponding to the recommended videos of each cluster in descending order to obtain the display order of the recommended videos corresponding to each cluster.

[0076] The recommended videos for each cluster are sorted from highest to lowest based on the closeness of the close friends who are discussing the videos. This results in the display order of the recommended videos for each cluster.

[0077] Step S35: Recommend the videos to the user for playback according to the display order of the recommended videos corresponding to each cluster.

[0078] Once the display order of the recommended videos for each cluster is obtained, they are recommended to the user for playback.

[0079] Based on the same inventive concept Figure 4 This is a schematic diagram of a video recommendation system based on big data, provided as an embodiment of the present invention. The video recommendation system based on big data includes:

[0080] The first acquisition module 41 is used to acquire user chat history;

[0081] The video output module 42 is used to input the user chat history into the video output model to determine multiple discussion videos in the user chat history and the discussion chat history corresponding to each discussion video;

[0082] Friend output module 43 is used to determine multiple close discussion friends and the degree of closeness between the multiple close discussion friends and the user based on multiple discussion videos in the user's chat history and the discussion chat history corresponding to each discussion video using a friend output model;

[0083] The second acquisition module 44 is used to acquire the video browsing history of each close friend and the user's video browsing history;

[0084] Module 45 is used to construct a graph structure, which includes multiple nodes and multiple edges between the nodes. The multiple nodes include a user node and multiple close discussion friend nodes. The user node is the central node. The multiple close discussion friend nodes establish multiple edges with the user node. The node feature of the user node is the user's video browsing history. The node feature of each close discussion friend node is the video browsing history of each close discussion friend. The feature of the multiple edges is the degree of closeness between the multiple close discussion friends and the user.

[0085] The recommended video determination module 46 is used to process the graph structure based on a graph neural network model to determine multiple recommended videos;

[0086] The recommendation module 47 is used to make recommendations to users based on the multiple recommended videos.

Claims

1. A method for video recommendation based on big data, characterized in that, include: Obtain user chat history; The user chat history is input into a video output model to determine multiple discussion videos in the user chat history and the discussion chat history corresponding to each discussion video. The video output model is a Transformer model. The input of the video output model is the user chat history, and the output of the video output model is multiple discussion videos in the user chat history and the discussion chat history corresponding to each discussion video. The video output model includes a chat history initial screening layer, a discussion video output layer, and a chat history determination layer. The chat history initial screening layer, the discussion video output layer, and the chat history determination layer all include a Transformer structure. The input of the chat history initial screening layer is the user chat history, and the output of the chat history initial screening layer is a preliminary selection of video-related discussion chat history. The input of the discussion video output layer is the preliminary selection of video-related discussion chat history, and the output of the discussion video output layer is multiple discussion videos. The input of the chat history determination layer is multiple discussion videos and the preliminary selection of video-related discussion chat history, and the output of the chat history determination layer is the discussion chat history corresponding to each discussion video. Based on multiple discussion videos in the user's chat history and the discussion chat history corresponding to each discussion video, a friend output model is used to determine multiple close discussion friends and the degree of closeness between the multiple close discussion friends and the user. Obtain the video browsing history of each close friend and the user's video browsing history; A graph structure is constructed, comprising multiple nodes and multiple edges between the nodes. The multiple nodes include a user node and multiple close discussion friend nodes. The user node is the central node. Each of the multiple close discussion friend nodes establishes multiple edges with the user node. The node feature of the user node is the user's video browsing history. The node feature of each close discussion friend node is the video browsing history of each close discussion friend. The feature of each edge is the degree of closeness between the multiple close discussion friends and the user. Multiple recommended videos are determined by processing the graph structure based on a graph neural network model; Recommending users based on the multiple recommended videos, wherein the recommendation based on the multiple recommended videos includes: K-means clustering is performed on the multiple recommended videos to obtain K clusters and the cluster center of each of the K clusters; The recommended video for each cluster is obtained by finding the recommended video closest to the cluster center of each cluster. Determine the closeness of the close discussion friends corresponding to the recommended video for each cluster; The display order of the recommended videos for each cluster is obtained by arranging the closeness of the close discussion friends corresponding to the recommended videos in descending order; The recommended videos for each cluster are recommended to the user for playback in the order they are displayed.

2. The video recommendation method based on big data as described in claim 1, characterized in that, The multiple recommended videos are selected from the video browsing history of each of the close friends.

3. A video recommendation system based on big data, characterized in that, include: The first acquisition module is used to acquire user chat history; A video output module is used to input the user chat history into a video output model to determine multiple discussion videos in the user chat history and the discussion chat history corresponding to each discussion video. The video output model is a Transformer model. The input of the video output model is the user chat history, and the output of the video output model is multiple discussion videos in the user chat history and the discussion chat history corresponding to each discussion video. The video output model includes a chat history initial screening layer, a discussion video output layer, and a chat history determination layer. The chat history initial screening layer, the discussion video output layer, and the chat history determination layer all include a Transformer structure. The input of the chat history initial screening layer is the user chat history, and the output of the chat history initial screening layer is a preliminary selection of video-related discussion chat history. The input of the discussion video output layer is the preliminary selection of video-related discussion chat history, and the output of the discussion video output layer is multiple discussion videos. The input of the chat history determination layer is multiple discussion videos and the preliminary selection of video-related discussion chat history, and the output of the chat history determination layer is the discussion chat history corresponding to each discussion video. The friend output module is used to determine multiple close discussion friends and the degree of closeness between the multiple close discussion friends and the user based on multiple discussion videos in the user's chat history and the discussion chat history corresponding to each discussion video using the friend output model. The second acquisition module is used to acquire the video browsing history of each close friend and the user's video browsing history; A construction module is used to construct a graph structure, which includes multiple nodes and multiple edges between the nodes. The multiple nodes include a user node and multiple close discussion friend nodes. The user node is the central node. The multiple close discussion friend nodes each establish multiple edges with the user node. The node feature of the user node is the user's video browsing history. The node feature of each close discussion friend node is the video browsing history of each close discussion friend. The feature of the multiple edges is the degree of closeness between the multiple close discussion friends and the user. The recommended video determination module is used to process the graph structure based on a graph neural network model to determine multiple recommended videos; The recommendation module is used to recommend users based on the multiple recommended videos. The recommendation module is also used to: perform K-means clustering on the multiple recommended videos to obtain K clusters and the cluster center of each of the K clusters. The recommended video for each cluster is obtained by finding the recommended video closest to the cluster center of each cluster. Determine the closeness of the close discussion friends corresponding to the recommended video for each cluster; The display order of the recommended videos for each cluster is obtained by arranging the closeness of the close discussion friends corresponding to the recommended videos in descending order; The recommended videos for each cluster are recommended to the user for playback in the order they are displayed.

4. The video recommendation system based on big data as described in claim 3, characterized in that, The multiple recommended videos are selected from the video browsing history of each of the close friends.

Images