Hybrid Recommendation Algorithms Based On Symbolic Data Analysis

In nowadays, the Internet has been able to providing basic service to users;meanwhile, it is gradually developing toward the direction of socialization, mobilityand personalization. Among them, SNS is the typical symbol of socialization;Mobility helps users achieve surfing online at anytime and anywhere; Personalizationmeans the Internet service can meet user’s personal needs. Recommender system,which is regarded as an important way to achieve personalization has gainedwidespread concern in the field of e-commerce and entertainment. Thecontent-based(CB) algorithms provides understandable results, while it often becomeoverspecialized and it’s only applied in domains where it is possible to describe itemby features. Collaborative filtering (CF) may provide unexpected recommendationsitems. However, it’s influenced by sparse data problem and cold start problem.In order to find solutions to the problems above, this paper exploited theadvantages of the two basic methods and proposed two hybrid approaches based oncontent-based and collaborative filtering. The first one is a combination of CB anduser-based CF; we constructed the user profile using symbolic modal-valued data andcompared two users by the dissimilarity measures for symbolic modal-valued data. Inaddition, the comparison between user’s demographic information is achieved byusing their age and gender. Final similarity is composed of the two parts mentioned.The second approach involves non-negative matrix factorization (NMF) in smoothingthe feature matrix of items. The comparison between items is achieved based on the―smoothed‖matrix. Both of the hybrid approaches can avoid the impact of the sparsedata, while NMF can effectively deal with the high dimension problem.We implemented the two hybrid approaches in comparison with classicuser-based CF and item-based CF; Experimental results show that the novelalgorithms provide higher efficiency in the section of similarity computation andsignificantly improve the performance of dealing with sparse data and new userproblem in recommendation quality.