Research On Recommendation Diversity For Microblog Users

As a social network, microblog is changing the way people get informationand communication deeply. More and more people are willing to express theirviews and feelings, publish message in microblog. However, as user in microblogfollow more and more people, they get too much microblog messages whichcontain a number of messages that user doesn’t like. This problem bother userswhen they reading microblog messages and reduce the user experience.In microblog user get message in chronological order, this paper focus onplacing the message which user likes on the top of the list, which can help improveuser experience.First, we use feature-based matrix factorization to recommend messages.matrix factorization is the best method in traditional recommender system, andapplied in microblog messages recommendation. Simultaneously, considering thecharacteristic of microblog, we propose several content and social feature to getbetter performance. This paper define the microblog recommendation as a problemof list ranking and we evaluate our result using MAP which is popular in searchengine evaluation. The experiment show that our method can raise the message thatuser like up, and prove that reccmendation can improve user experience.Using recommendation can lead a problem that a large number of message inthe list are similar. Although user may like these message, it’s hard for user to getdiverse information, In the long term, it will hurt user’s experience. To solve thisproblem, we propose method to measure the diversity of a list and wish that thesimilarity of the messages of the list be low.We propose three types of feature to measure the similarity of messages anduse incremental single-pass and cluster ensemble to cluster mircoblog messages.We compare the clustering result with user annotation result and use NMI and B-cubed index to measure the performance. Result show that our feature are usefulwhen computing the similarity of messages. Based on the similarity of messages, we use greedy algorithm to diversify thelist and using online user annotation to verify that the diverse list get better userexperience than normal recommendation list.