Research On Service Recommendation Algorithms Combining Multivariate Information

With the vigorous development of the Internet and big data technology,the "information overload" problem has spawned a large number of personalized service recommendation systems.Because the collaborative filtering algorithm has high recommendation accuracy and is easy to implement in the engineering application field,it has been widely used in the recommendation field.However,there are several challenges for the collaborative filtering algorithms,such as data sparsity,cold start and scalability etc.Only utilizing past activities of users to make recommendations can not effectively solve the inherent problems existed in recommender systems.With the rapid growth of network user groups and the advancement of information technology,a large amount of multivariate information has been derived,such as demographic information,commodity information and social network information of users.Multivariate information has brought opportunities for optimization of algorithms.How to use the rich multivariate information to solve the problems in the collaborative filtering algorithm has become a research hotspot for personalized recommendation.In this thesis,we use the missing data imputation of matrix,user-based collaborative filtering recommendation and K-Means clustering as the research basis to analyze the shortcomings of the existing improved algorithms,and realize multivariate information fusion at each stage of the recommendation algorithm.The specific contents are as follows:Aiming at the problem named "Harry Potter" in the recommendation system,an improved Pearson similarity formula is designed.In the similarity calculation process,by decreasing the weight of popular product to improve accuracy of algorithm;according to the problem of overestimation,traditional improved similarity computation methods using Jaccard coefficient make corrections,but it's not carefully.Thus,this paper proposes a new method to optimize the relation between local similarity and global similarity.The new method keeps the simplicity and efficiency without additional complexity.Experimental results show that the new method outperform traditional methods or common significance weighting methods on the prediction of ratings.To deal with user cold-start,the thesis introduces user demography and modifies similarity calculation combined with both demography and ratings.The improved algorithm can dynamically adjust the proportion of the two according to different situations of different users.When a new user comes to the recommender system for the first time,optimized similarity calculation can help find nearest neighbors of the new user based on demographic similarities and then make recommendations.Aiming at data sparsity that is likely to affect clustering and recommendation results,the thesis prefills the rating matrix based on the similarity of item attributes,so as to effectively alleviate data sparsity.Compared with the traditional average and mode filling methods,the filling of missing values based on the similarity of item attributes fully can not only alleviates the problem of matrix sparsity,but also can take into account the personalized information of users.Furthermore,in order to solve the problem of poor real-time,K-Means is applied to cluster users offline.Using combined recommendation strategy can effectively improve the efficiency of finding nearest neighbors and enhance the real-time of the recommender system.Finally,the thesis verifies the proposed algorithm using Movile Lens₁M and Movie Lens₁00K.From experimental results,we can see that the two improved collaborative filtering methods have obvious improvements on recommendation performances.