Research On Recommendation Alogorithm Of Clustering-based Low Rank Matrix Completion

With the development of Internet and big data,useful information acquisition for people is becoming more and more difficult.As one of the tools to effectively alleviate the information overload,the recommender system not only provides users with quick access to information of interest,but also has great commercial value.However,the numerous user,item and user rating data in the system become one of the key factors leading to the low performance of the recommendation algorithm.To resolve above mentioned challenges,this thesis proposes two similar user partition schemes,and then proposes a low rank matrix completion recommender algorithm based on spectral clustering.The main research work of this thesis is as follows:(1)A method of normalized rating vector is proposed to track the user's preference record,which reduces the impact of rating sparseness on the similarity calculation between users.Two similar user classification mechanisms are proposed: a location-sensitive hash function classification algorithm and a graph theory-based spectral clustering algorithm,which partition the original user item matrix into multiple sub-matrices.(2)A low rank matrix completion algorithm based on spectral clustering is proposed.Based on the spectral clustering user classification method,in order to further improve user's rating rate,reduce the size of the sub-matrix,prune the columns unrelated to the target user.Then a joint optimization model is proposed,parameters include classification number and pruning rate are determined by optimizing the target loss function,which achieves the algorithm to optimal recommended performance.Finally the matrix completion algorithm is used in the submatrix.Experimental result indicate that our proposed two classification algorithms proposed in this thesis can effectively improve the partition accuracy of similar users.Based on the spectral clustering classification algorithm,the proposed low rank matrix completion joint optimization algorithm can effectively improve the prediction accuracy of the interested items to users.