Research On Time-sync Comments Based Video Recommendation System

With the rapid development of the Internet,technologies such as multimedia processing,network transmission,and video data management have developed rapidly,and people have access to more content from various video websites.The dazzling information is filled with people's eyes and brains.In an age when vision and hearing are so prosperous,the recommendation system as a solution to information overload has gradually attracted more and more attention in academia and industry.With the further development of the recommendation system,scholars began to focus not only on the interaction record between the user and the video,but also on the attributes of the user and the video itself.Due to the development of machine learning,especially deep learning,many methods have been applied in the research to combine video images,video comments and other information with recommendation systems.Time-sync comment is a form of commentary that has become more and more popular on an Internet platform in recent years.It enables users to express real-time opinions or comments on current video clips,and the research on time-sync comment is still in its infancy at home and abroad.It mainly includes two application fields: video keyword extraction and video highlight segment extraction.Like pictures and comments,the time-sync comment contains a lot of information about user preferences and videos features,so extracting semantic information from it is an important way to improve the effectiveness of the recommendation system.The time-sync comment has many characteristics compared to the traditional comment,including real-time,spontaneity,diversity,richness and interactivity.On the one hand,the users express their interest in the current video content through the time-sync comments,on the other hand the topics in different time-sync comments constitute an overview of the video content,so from the time-sync comments we can discover the interest of each user and the topic distribution of the video.The user and video features extracted from it can help the recommendation model perform better.However,the arbitrariness,brevity and colloquialism of the time-sync comment make it mixed with large quantity of noise,which forms obstacles to semantic extraction.Therefore,how to make full use of features in time-sync comment and apply them in the recommendation system is the main research content of this paper.This paper proposes a recommendation system based on the time-sync comment,analyzes the user and item features,and makes recommendation based on these features.This paper mainly contains the following research work:(1)Collecting the time-sync comments from the video website,aggregating all the user's comments into a user data set,aggregating all the comments contained in the video into a video data set,and then processing the item and video data separately.Analyzing statistics of data set and the recommendation performance based on different data characteristics such as data sparsity and video type.(2)In order to balance the performance of the recommender model and the training time complexity,this paper proposes two recommendation algorithms based on time-sync comments:the personalized topic recommendation model and the personalized recurrent recommendation model.The former uses the topic model to process the time-sync comments to extract user and video features,and designs a multi-layered perceptual neural network in the upper layer to modify features using interaction data between users and videos to improve recommendation performance;the latter uses word vector to process time-sync comments and designs a recurrent neural network based on the attention mechanism to extract user and video features for recommendation.(3)The proposed algorithm is compared with the classical and the deep learning-based recommendation algorithm on the actual data set to analyze the recommendation performance.(4)Design and implement a prototype system based on the time-sync comments.The main functions of the system include user registration,login,watching video and time-sync comments,sending time-sync comments,etc.The system collects user's viewing and comment records,and makes personalized recommendation for them.This paper firstly introduces the background and significance of the recommendation system based on the time-sync comments,analyzes the current situation and existing problems and puts forward the technical route of this research.Then two models proposed in this paper are described in detail,including models parameters,structure,training methods and experimental performance.Finally,a prototype of recommendation system was designed and implemented to show the real application scenarios of the model.