Research On Collaborative Filtering Recommendation Methods Oriented User Data Characteristic

In the Internet era,the explosive rise of information and unprecedented data size have greatly exceeded the receivers' receiving and handling capability.Accordingly it has become a necessity to capture important and useful information from the massive and complex data.The problem of information overload calls for a more efficient information filtering system and this explains the birth and rise of the recommendation system.Important progress has been made since 1990 s in terms of theory,methodology and application of recommendation system but challenges still exist in the areas such as sparsity and long tail of data,users' behavior pattern mining,interpretability and social recommendation.In addition,with the increasingly rapid development of Internet and information technology,users' scale and the number of items dramatically surge and the data are featured by sparsity and long tail more than ever.To address these challenges,the author attempts to solve the issues such as the incapability of traditional similarity measure,the nearest neighbor recommendation's sensitivity to data sparsity,interpretability caused by the long tail data and the social information integration by introducing the innovative collaborative filtering recommendation method.The research results are concluded as follows:To address the incapability of traditional similarity measure for the sparse users' rating matrix data,the author proposes nearest neighbour recommendation method basing on the users' extreme rating behaviors.To address the incapability or inaccuracy of traditional similarity measure caused by the scarcity of common rating items in the sparse data,the author makes a deep analysis on the users' common extreme behavior and establishes an index among the users' extreme rating behaviors.The index is then integrated as weight with the traditional correlation coefficients to establish the similarity measurement principle featured by the fusion of non-liner and liner correlations.The experimental results show that the nearest neighbour recommendation method more elaborately describes the users' similarity and improves recommendation precision.To overcome the nearest neighbour recommendation's sensitivity to data sparsity,the author proposes probability matrix factorization method fusing neighbour information.Due to the limited number of common rating items and amount of neighbour information for nearest neighbour recommendation,the recommendation robustness is affected.The author attempts to establish neighbour matrix basing on the similarity of extreme rating behaviors to moderate the instability in capturing the neighbour information.The adaptive probability matrix factorization method can fit the rating matrix and show robustness in prediction.So the collaborative filtering recommendation method integrating extreme rating behaviors,neighbour relationship and probability matrix factorization model is built up.The experimental results show that the probability matrix factorization method is robust in prediction accuracy.To handle the interpretability of long tail recommendation,a threefactor probabilistic graphical model is proposed.To meet the recommendation system and users' need for interpretable item recommendation in reality,the author proposes a three-factor probabilistic graphical recommendation method basing on customers' activity level,the items' unpopularity and customer-item preference level.In this method,the advantage of probabilistic graphical model in interpreting cause-result relationship is utilized and recommendation accuracy and design novelty in the algorithm can be ensured in this study.The results of experiments indicate that with the advantage of interpretability,the three-factor probabilistic graphical recommendation method can ensure prediction accuracy and perform better in providing novel recommendations.The author further proposes long tail recommendation method integrating social network information.Inspired by the basic fact that the users tend to trust their friends' recommendation,the author attempts to integrate the users' social network information and their rating matrix information,share the users' latent feature vector and include the friends' recommendation information into the long tail recommendation framework in the probabilistic graphical model so as to provide a probabilistic graphical recommendation method involving users' social network information.Experiments results indicate that the method ensures prediction accuracy and provides more satisfactory results in long tail recommendation.To conclude,to address the challenges brought about by the sparse and long tail data,this study proposes collaborative filtering recommendation methods incorporating the techniques such as similarity measure,nearest neighbour recommendation,probability matrix factorization and probabilistic graphical to cope with the information overload problem in the Internet era.The research results enrich and develop the extant recommendation model and methodology and are expected to have application prospects in the real recommendation scenarios in E-commerce websites and social network media.