A Tensor Decomposition Method Based On Stochastic Gradient Descent And Its Application In Video Tag Recommendation

In video sites such as movielens and Douban,users label videos to form video description documents,and also form a user's description document.The video tag recommendation system analyzes the user's preferences based on these information and recommends suitable tags for the user.With the explosive growth of network data,video resources and corresponding video tags are presented to users more and more.How to accurately and quickly provide users with corresponding tags to meet the personalized needs of users has become an urgent problem to be solved.The three-dimensional tensor represented by “user,video,tag” can well represent the relationship among the three main elements in the tag recommendation system,and further use tensor decomposition to find the semantic relationship among them and recommend relatively accurate tags for users.However,the data distortion problem existing in the tensor construction and the data sparsity in the tensor decomposition process all affect the accuracy of the recommendation.In view of the above problems,this paper studies the video tag recommendation algorithm based on tensor decomposition from three aspects and puts forward some innovative points:1)In the tensor construction phase,aiming at the "0/1" filling method,this paper proposes a tensor construction method PMUS(Penalty Mechanism-User Score)which combines the penalty mechanism and the user rating.The weights in the resources and the differences in interest of different users can effectively eliminate the negative influence of popular tags.The experimental results show that the PMUS tensor construction method can effectively improve the accuracy of label recommendation.2)In the tensor decomposition stage,based on the data sparseness problem,a stochastic gradient descent tensor decomposition algorithm HOSGD(High Order Stochastic Gradient Descent)is proposed.In the decomposition stage of the tensor expansion matrix,a stochastic gradient descent method(SGD)is used.The matrix is processed.Experimental results show that the HOSGD tensor decomposition algorithm can handle sparse data well.3)After the tensor decomposition is completed,further consider the relationship between friends,propose a tensor correction method that integrates friends,and combine the similarity of friends to correct the results after tensor decomposition to obtain a fused relationship with friends.Recommended list.The experimental results show that the tensor correction method for merging friends can effectively correct the error of the original tensor result and improve the accuracy of the recommendation.