Increasing Serendipity of Recommender System with Ranking Topic Model

There are thousands of academic paper published each year, it is quite hard for researchers who enters a new field to discover relevant paper and novel paper to read, which we characterize as choice overload problem. Recommender system can help to alleviate the problem, but recommender system suffers from the intention gap problem which is the incapability of the system to accurately guess users' intentions. We proposed a ranking topic model based semantic recommendation framework which helps to introduce serendipity to the system. First, the proposed ranking topic model reorders learnt topic distributions according to users' intentions. Then, learnt ordered topics are used as features to rank papers in the library according to the relevancy to user query. At the same time, ranked topics also provide novelty to the results. Since there is little work on how to evaluate the serendipity degree of recommender system, we proposed two measure to evaluate this metric. We performed empirical experiments to test the efficiency of proposed framework with state-of-the-art counterparts, the comparison results revealed the superiority of our proposed algorithms. In the end, we illustrated our algorithms with an example and pointed out future research directions.