Research On Video Recommendation Based On Hypergraph And Fusion Of Multi-features

With the rapid development of the Internet technology,more and more users prefer to watch online videos on the Internet.However,with the rapid growth of the number of videos on various websites,the hazards of information overload also follow.When facing the huge amount of video resources on the Internet,users often find it difficult to quickly find the videos they are interested in.In order to filter out videos that meet user interests from a large number of videos,recommendation system is indispensable.Unlike other products that often have clear attributes and keywords,it is difficult for online videos to directly determine their characteristics through clear attributes and keywords.Traditional video recommendation methods often ignore the high-order correlations between users,and do not combine features of different granularities well,which will lead to information loss or information cocoon problems,which makes the recommendation performance unsatisfactory.In response to the above problems,this thesis mainly conducts research on video recommendation methods based on hypergraph and multi-feature fusion.The main work is as follows:First,in the recall phase,this thesis uses coarse-grained modeling of video categories and tags to reduce the problem of information cocoon rooms,and regards the categories and tags as keywords to extract word-level semantics to obtain word vectors to represent video information.According to the attention mechanism,different videos in the user’s historical interaction records are assigned different importance,and the video candidate set is obtained by comparing the similarity between the user feature vector and the target video feature vector,so as to realize the recall of the video recommendation.Then,in the ranking phase,Multi-Feature Video Recommendation based on Hypergraph Convolution is proposed.This method performs feature representation based on the user-video-tag relationship,and uses a multi-layer perceptron to reduce the dimensionality of high-dimensional sparse vectors,and then treats the user as a node to construct the hypergraph.According to the obtained supergraph structure,the supergraph is based on the spectral method.Graph convolution,by aggregating the information on the nodes on the super edge and then on the nodes to transfer to update the features.According to the attention mechanism,the user’s historical interaction record is modeled to extract the user-side item representation feature,and the user-side feature is fused and expressed,thereby sorting the video candidate set obtained before to obtain the video recommendation list.Finally,based on the theoretical methods mentioned above,this thesis conducted simulation experiments and analysis on real datasets to demonstrate its feasibility and effectiveness.In addition,in order to realize the combination of theoretical methods and practical applications,this thesis designs and implements a prototype system,including demand analysis,overall design,specific design and operation flow,etc.,and further demonstrates the performance of this method.