Research On Collaborative Filtering Algorithm Based On Association Rules Of Tags

There are three stages in the process of the development of information retrieval tools: classification navigation technology,search engines,and recommendation system.Recommendation system and some other related technologies have been in our life,it has been used on the sites of video,music,social networking,and even the news we browse every day.There are many defects in traditional algorithms of collaborative filtering recommendation,such as the problems of sparsity,cold booting,expansibility,user's multiple interests,and so on.With the development of Web2.0,the tag elements added in the social tagging system provide the users a new way to express the subjective feelings of the items.Tags not only can reflect the users' views and interests of the items,but also give a really accurate description of the contents of the items.User-generated content(UGC)has a very important significance on carring out social interests of the users on the Internet.In this paper we propose a collaborative filtering algorithm that based on association rules of tags which introduces a user-defined label content.Firstly,we applied the item-based collaborative filtering into the usrs' rating matrix filling,it is an effective method to solve the problems of sparsity in the algorithms of traditional collaborative filtering.And we introduce the user's attention matrix,improve the similarity based on similarity of the user's scores and the attention of user's degrees in order to improve the similarity of the users.Here we introduce the idea of frequent item sets in the association rules of Apriori to train the suitable minimum supported threshold,find frequent item sets,get the points of users' interest,and then traverse the set of users in a backward way,clustering the users according to the points of user' interest.After this,we calculate the similarity between users using the improved method based on user similarity described above,find the user's neighbor set,and then we can predict the rating of the item provided by the users,and after all above we recommend the items to users.We test it use a series of comparative experiments based on the Movie Lens film score data set.Experiments show that the proposed method can effectively reduce the impact of the sparse score matrix,and improve the prediction accuracy.