Topic Collaborative Filtering And Its Application

Researchers nowadays have more access to research archives then ever due to therapid development of communication tools and Internet. All kinds of research data can betransformed into digital files and spread through network in a short time. However thehigh speed growth of research archives brings the side effect which makes finding arelevant paper become more difficult. Then the appearance of online scientificcommunities such as CiteULike helps a lot. These sites allow the registered user toconstruct a private library which contains the paper she likes, and she also can share itwith other users in the community. Managing lots of paper becomes more organized andeffective. However, it does not solve the fundamental problem we mentioned above, theusers still need to spend lots of time finding an interesting paper.The major cause is that our understanding about the user and articles is poor.Collaborative filtering is an effective way to model and understand users. Based on thepast behaviors of users, it utilizes techniques from machine learning and data mining toanalyze historical data, in order to provide relevant recommendations. However,collaborative filtering suffers the problem of ‘cold start’. On the other hand, probabilistictopic models can help modeling articles in an elegant manner, which provides strongsupport in analyzing the content of documents. Nevertheless, topic models do not takeuser information into consideration.A probabilistic model which combines both traditional collaborative filtering andtopic models can be an appropriate solution to solve this problem. It provides ainterpretable framework to model both users and articles: on one hand, it deals with userinformation to enhance our understanding about users’ interests; on the other hand, itanalyzes articles’ content to improve our comprehension about the documents. As it takesadvantages from both collaborative filtering and content analysis, it can makerecommendation for users intelligently. We study this model on the dataset fromCiteULike, the result proves that comparing with traditional algorithms, this model canimprove the performance significantly.