Research On Collaborative Filtering Recommendation Algorithms For Sparse Data

With the advent of the era of intelligent Internet,people urgently need some technical means to automatically and quickly find the data that meet the interests of users from massive data,and personalized recommendation technology is born.Collaborative filtering(CF)algorithm is one of the most mature recommended technologies.In the face of the growing number of large-scale users and items,due to the limitation of user ratings and the increase of new users and items,CF algorithm still has some problems that need to be solved urgently.In view of the defects of CF algorithm and the shortcomings of existing problem solving methods,this thesis deeply study of CF algorithm,and puts forward improvement schemes and measures to achieve more accurate recommendation.CF algorithm mines users' preferences through historical records of different users,and then divides users into groups based on different user preferences by using similarity,and recommends items that are liked by similar neighbors.Since most of the CF algorithm is based on a single measure similarity score matrix,while ignoring the distribution characteristics of different item scoring vectors and the influence of the number of users with too little common score on similarity calculation.At the same time,it does not consider the influence of multi-feature fusion on user interest,such as the influence of comment text,the consistency of score and comment,and the fine-grained emotional score in text.In view of this,this thesis focuses on the data sparsity problem faced by the algorithm,and makes a thorough study on similarity and multi-feature fusion.Specifically,this thesis mainly does the following work:1.Research on mutual information item collaborative filtering recommendation algorithm.Similarity calculation is the basis of CF algorithm,and data sparsity often leads to inaccurate similarity calculation,which affects the accuracy of recommendation.Aiming at the shortcomings of the traditional similarity method,this thesis improves the similarity method from the two factors of data distribution characteristics and data sparsity.The algorithm takes full account of the distribution of scoring vectors of different item,uses mutual information to mine the correlation between items,and then introduces smoothing coefficient to alleviate the impact of the number of users with too few common scoring on similarity calculation.The experimental results show that the improved algorithm has certain advantages over other methods in scoring and prediction accuracy.2.Research on the Collaborative filtering recommendation algorithm for coarse-grained affective analysis.Comment text is an important means to alleviate the problem of data sparsity.How to accurately mine users' interest in different fine-grained aspects of the item to improve system accuracy is a hot topic in the field of recommendation.This thesis makes full use of multi-feature fusion such as comment text's own influence,emotional expression score and consistency of score and comment to mine user's interest value in different aspects of the item,to construct user-item-aspect interest matrix,and then calculate user-to-user similarity in fine-grained,then fuse user-to-user similarity in coarse-grained score matrix,and finally recommend.Relevant research has been carried out on the crawled Hotel comment data set of TripAdvisor website.The experimental results show that the problem of data sparsity has been further alleviated by fusing fine-grained emotional information in the comment text,and the performance of the recommendation algorithm has also been improved.