The Research Of Collaborative Filtering Recommendation By Integrating Trust And Similarity

With the rapid development of information technology and the Internet,it is difficult for users to find useful information quickly in the massive data,which lead to the problem of information overload.As a tool to effectively deal with the information overload,the personalized recommendation technology has been widely concerned in all walks of life.For the advantages of simple principle,easy implementation and strong interpretability,the collaborative filtering recommendation algorithm has become the most widely used personalized recommendation algorithm.The algorithm mines the interest preferences based on historical information,and the key step is to calculate the similarity between users.However,due to the sparseness of data and malicious users' rating,as well as the its own defects of collaborative filtering algorithms,the reliability of the similarity calculation is poor,which result in the unsatisfactory recommendation.How to improve the problems that recommender system face is a hot topic in personalized recommendation research.Further research is conducted with the above issues in this thesis.The main work presents as follows:1.The results of the similarity obtained by the existing methods are not accurate enough when the rating matrix is sparse,which result the large deviation when predict the missing rates.Therefore,an optimized algorithm of similarity calculation is proposed in this thesis.Firstly,the algorithm calculates the coincidence degree of the users' common rating items and reduces the deviation of the similarity calculation caused by the less common items.Secondly,the contribution of the variance is improved,and reduce the impact of the larger deviation on common items.And then,by the dynamic adjustment of the proportion of the common items,the shrinkage factor and the reliability of the similarity calculation are improved.Finally,the penalty function is introduced to reduce the impact that caused by popular items,which improves the reliability of the similarity calculation and the accuracy of the recommendation.2.For the possible problem of shilling attack in the collaborative filtering recommendation,the trust relationship is explored.The trust model is constructed to calculate the trust between users and combine it with the improved similarity.Integrating the two factors of the trust degree and the similarity to select neighbor users,which improve the robustness of the collaborative filtering recommendation algorithm.The improved similarity calculation method and the trust model proposed in this thesis are simulated on the two public datasets.The results show that the improved similarity calculation could predict users' rating more accurate in the sparse data,and the trust model could improve the robustness when the system was rated by the malicious rating users.