A Survey Of Entropy Based Personalized Recommendation

Recommender system is being an indispensable part of E-commerce. It collect-s user behavior data and analyzes correlations between items to provide appropriate recommendations for the users who visit the E-commerce site. When the users visit item pages, recommender system generates recommendations for the visited items by item-based collaborative filtering according to the similarities between items. How-ever, such recommendations cause the users who have various preferences on items receiving monotonous recommendations due to lack of personalization.Personalized technology uses statistics, probability theory, ontology and informa-tion retrieval to analyze user interests; measures related degrees of interests and store them into user profiles; re-ranks the order according to user profiles. In this paper, two available techniques-TF and TF-IDF are used to measure user preferences on attributes, so as to personalize the recommendations generated by item-based collab-orative filtering (CF). In details, TF measures user preferences on attributes by the statistics of user visits on attributes, while TF-IDF introduces the relevances between items and attributes to refine user preferences on the basis of TF.Besides that, considering the priorities of attribute comparing process, using at-tribute entropy to facilitate existing personalized methods is proposed in this paper. Attribute entropy is a kind of diversity entropy, and is calculated by attribute class and attribute distribution in corresponding class. Attribute entropy is transformed into attribute entropy weight according to users’reasonable behaviors and impulsive be-haviors. Both TF and TF-IDF could be improved by attribute entropy weight. The experimental results show that the improvements are effective and significant.Due to public data set-MovieLens not containing the detailed attribute informa-tion, the data used in experiments are collected from a real E-commerce site, and the time span is about6months. In this paper, temporal experiments are used to evaluate the effects of personalization. The results show that personalized technology could improve item-based recommendation lists in terms of accuracy, novelty as well as cov-erage to10%,25%and25%respectively. And attribute entropy based improvement could increase these metrics further to another10%.