Research On Dynamic Collaborative Filtering Algorithm Based On Latent Semantic Model

The recommendation system collects user's behavior history data,constructs related models to predict user needs and interests,filters useful information for users,and provides solutions for the “information overload” problem in the current big data era.Collaborative filtering is an algorithm that is widely used in recommendation systems.It requires only a small amount of historical scoring data between "user projects" to efficiently build recommendation systems that anticipate the needs of users.However,with the expansion of data size,the increase in the richness of data types and the diversity of application environments.Data sparsity,scalability,real-time and so on are the main problems encountered in current collaborative filtering algorithms.In this paper,the problem of data sparsity and real-time in collaborative filtering algorithm is studied,and the collaborative filtering algorithm based on implicit semantic model is improved.The main work is reflected in the following two aspects:First,real-time performance has always been a major problem in collaborative filtering algorithms.The user's preference for the project will change over time,and the project itself will show different concepts at different time periods.Therefore,dynamic modeling of time is indispensable for designing a recommendation system or general user preference model.This paper improves on the basis of the traditional SVD++ algorithm,and models the changing rules of the score data over time.It integrates the user preference feature factor time information,user time offset and project time offset,and on this basis also adds User characteristics information,a new improved algorithm is proposed: SpecialTSVD++ algorithm.The improved algorithm in this paper strengthens the connection between data and time,reflects the dynamic changes of the recommendation results over time,and improves the accuracy of the prediction score.Secondly,aiming at the data sparsity problem of collaborative filtering algorithms,this paper proposes the ESpecialTSVD++ algorithm combined with the density-based active learning strategy and SpecialTSVD++ algorithm.This algorithm reduces the sparsity of the scoring matrix by automatically adding “score” to existing users.Sex.The method divides the original matrix into dense sub-matrices.The scoring data of the sub-matrix is provided by users with more scores.Then the default values in the submatrix are predicted and filled using the Special TSVD++ algorithm to obtain a complete matrix.Finally,the completely filled sub-matrix is integrated into the original matrix,which increases the number of known scores in the original matrix,making the original data more abundant and the recommendation accuracy improved.Further,this paper proposes a multi-layer model based on the original single-layer ESpecialTSVD++ model.The multi-layered ESpecialTSVD++ model solves the problem that the single-layer model has large deviations from the original model data through iterative training.