Research On Movie Recommendation Technology For Sparse Rating Data

With the continuous development of the Internet,the data on the Internet has exploded.Although massive amounts of information have brought great convenience to people in all aspects of life and enriched people's demand for information,the huge amount of data has increased people's access the degree of difficulty in having useful information about itself has caused an overload of information.The same is true for movie information.As the number of movies continues to increase,it becomes more difficult for users to find their favorite movies.How to find a movie that every user likes from a wide variety of huge amounts of movie data has become a problem that must be studied.As one of the effective methods to deal with such problems,the movie recommendation system has become the focus of research by movie recommenders and researchers.The core of the movie recommendation system is the movie recommendation algorithm.The research of the movie recommendation algorithm focuses on two aspects,which are top-N list recommendation and rating prediction.Top-N list recommendation research focuses on how to generate satisfactory movie recommendation lists for users,while rating prediction focuses more on predicting user ratings for certain movies.Matrix factorization recommendation algorithm is widely used in movie top-N list recommendation because of its easy implementation and low time complexity.However,the traditional matrix factorization recommendation algorithm has the following problems: when the rating data is too sparse,the recommendation performance of the matrix factorization recommendation algorithm will decrease;when the rating matrix is constructed,the explicit rating and implicit feedback cannot be reasonably combined.The memory-based collaborative filtering algorithm is widely used in movie rating prediction because of its simple ideas and easy implementation.However,the accuracy of the collaborative filtering algorithm's rating prediction will decrease when the rating data is too sparse or the calculation method of the similarity is not selected properly.This paper studies the problems existing in these two directions.The main work is as follows:1.Aiming at the impact of the sparseness of rating data on the matrix factorization recommendation algorithm,based on the deep matrix factorization recommendation model,an asymmetric deep matrix factorization recommendation model is proposed.The model is a new matrix factorization model with an asymmetric neural network architecture.It fully considers the effect of explicit rating and implicit feedback on recommendation results.Combining explicit rating and implicit feedback,a model is proposed.New interaction matrix filling method to improve the utilization of sparse rating data and alleviate the impact of rating data sparsity on model recommendation accuracy;Considering the difference in the number of users and movies,an asymmetric neural network is used to learn the interaction matrix to obtain more reasonable parameters.On the MovieLens-100 K and MovieLens-1M datasets,the proposed asymmetric deep matrix factorization recommendation model is compared with related research results,and HR and NDCG are used as evaluation indicators.The experimental results show that the proposed recommendation model can be more accurate to recommend the top-N list of movies.2.Aiming at the influence of rating data sparsity and similarity calculation method on collaborative filtering algorithm,a collaborative filtering algorithm based on Gaussian mixture model and improved Jaccard similarity is proposed.Firstly,a new method for constructing an interaction matrix is proposed.The Gaussian mixture model is used to cluster the rating data.According to the different clustering objects,the algorithm can be split into algorithm one for clustering user to build an interaction matrix and algorithm two for clustering movies to build an interaction matrix.Based on the clustering results,the corresponding users and movies are labeled with the cluster labels of their respective clusters,and a new user-movie interaction matrix is constructed according to the users,movies,and the cluster labels to which they belong,which alleviate the impact of rating data sparsity on the rating prediction effect of collaborative filtering algorithm;Secondly,the algorithm analyzes the shortcomings of Jaccard similarity.Jaccard similarity only considers whether the user rated the movie when calculating the similarity,and ignores the impact of the rating on the similarity measure.Therefore,the Triangle similarity and Jaccard similarity are combined.A new similarity calculation method is proposed to eliminate this deficiency.In the MovieLens-100 K,MovieLens-1M and Yahoo! Webscope R4 movie datasets,the proposed algorithm is compared with four existing collaborative filtering-based recommendation algorithms,using MAE as the evaluation indicator.Experimental results show that,compared with the four existing collaborative filteringbased recommendation algorithms,the proposed algorithm obtains more accurate rating prediction results.