Research On Collaborative Filtering Recommendation Model For Sparse Data

With the development of network platforms,the recommendation system has received extensive attention as a bridge between users and products.Collaborative filtering recommendation model is very popular in commercial recommendation.It mainly analyzes the interaction data between users and products,and establishes a recommendation model to predict users' preference for products.In most platforms,due to the large variety of products available for selection but the limited purchasing power of consumers,users' purchase or evaluation data for products only account for a small part of the number of products,resulting in very sparse interaction data between users and products.In order to improve the quality of recommendations,many scholars solve the problems caused by sparse data from various aspects.For the problem of sparse data,this paper proposes a matrix decomposition model based on the credibility of user ratings to alleviate the problem of sparse data.The matrix decomposition model decomposes the user and item rating matrix into two matrices:user and item,and predicts the user's rating of the item by training the model.This paper adds a global bias to the traditional matrix decomposition prediction formula,that is,the user's average score,which can improve the impact of individual singular score values on the prediction results,and also adds the user's implicit for other items in the user's implicit vector part.Scoring increases the utilization rate of user ratings and improves the credibility of user ratings.Experiments show that the model has higher prediction scoring accuracy.This paper also proposes a collaborative filtering recommendation model to improve user similarity.The model starts from the interaction data between users and items,calculates the local similarity between users for the common scoring item information between users,and the similarity calculation process uses a nonlinear model,and adds singularity factors to adjust the weights;For the overall rating data,the global similarity is calculated,the user's rating habits are added,and the two similarities are fused to obtain the recommended model.The model can effectively use sparse data,and experiments show that the model has better prediction accuracy.Through experimental comparison,it can be found that the two models have corresponding advantages.The collaborative filtering recommendation model combining local and global similarity uses all the scoring data to provide more accurate and detailed personalized recommendations,with good prediction scoring accuracy and recommendation accuracy.However,due to the complicated calculation process,the model has poor time accuracy.Based on the matrix decomposition model of user rating credibility,the implicit information of user ratings and global bias are added to improve the problem of uneven user ratings,with high calculation efficiency and good prediction accuracy.For different needs,two models can be selected.