Research On Hybrid Recommendation Algorithms Based On Matrix Decomposition

As a valuable information mining system,recommendation system plays an important role in supporting many websites,especially some shopping websites or social networking sites.Matrix decomposition is a very important branch of recommendation algorithm.Its greatest advantage is to solve the problem of sparse matrix.However,any excellent algorithm has certain limitations,so this paper studies a hybrid algorithm composed of multiple algorithms.A label-based matrix factorization recommendation algorithm is proposed.This algorithm adds label features to the traditional Latent Factorization Model,and decomposes the original score matrix into two matrices,user matrix and item matrix.On the basis of the classification of the features of the historical score data,the label type is added as a new feature value.Experiments show that adding tags to the model can effectively improve the recommendation accuracy.A probability-based non-negative matrix factorization recommendation algorithm is proposed.The algorithm randomly initializes two matrices to describe users and items respectively;uses probability graph model to iteratively optimize two matrices,and effectively prevents over-fitting of the algorithm by simplifying complex models;adds user bias matrix and item bias matrix to participate in the iteration calculation of user matrix and item matrix,thereby improving the accuracy of recommendation.Experiments show that the algorithm can significantly optimize the recommendation effect from various indicators.When the user's historical behavior is small or empty,the above algorithms can not achieve good results.Therefore,a recommendation algorithm is proposed to solve the cold start problem of the system.Combining the existing information of users,items and so on to optimize the model and predict.Experiments show that the algorithm can effectively improve the cold start problem of the system.A personalized recommendation system engine is designed.The engine preprocesses the data to prepare for the hybrid recommendation algorithm,initializes the models of each algorithm,then chooses the algorithm according to different types of users,finally returns the recommendation results to users,and can feedback the real-time value of the recommendation index of the system at each time,such as MAE.