The Paper Recommender System Based On Ranking Topic Model

We introduce a paper recommender system based on ranking topic model. This method not only ensures the accuracy of recommendation, but also improves the serendipity of rec-ommendation, and meets the potential needs of users. In this paper, we convert articles into different topics by topic model, the similarities of articles are decided by the relationships of topics, the crowdsourcing is proposed to measure the important of topics. We apply the paper recommendation system in mobile client by the WeChat public platform, collecting the user’s scores of serendipity, the scores are feedback to system and the weights of topics are then changed, ultimately, we get a list of topics order by the user’s feeling. Comparison results with other unsorted topic model and no feedback show that the ranking topic model can improve the serendipity of recommendation system.The paper consists of three aspects:1.The paper recommender system based on topic model.The LDA topic model can be used to train corpus. The paper is seen as many topics. We can recommend some interested papers to users according to the similarity of different topics.2.Collecting the users’ feedback by crowdsourcing. Two kinds of feedback strategies are used here:single-level feedback based on wechat users and multi-level feedback based on expert ratings. The users’ ratings can be accessed easily by wechat public platform to change the weights of different topics. The users’ feedback is useful to improve the serendipity of recommendation. However, the information of users’ feedback may be inaccuracy and random, then we take the method of multi-level feedback to proving the effect of feedback to recommendation. We choose the multi-level feedback of specific users’ ratings to change the weights of different topics.3.The serendipity of recommendation results. According to the model of evaluating, we present an assessment method to evaluate the serendipity of recommendation results. The serendipity is divided into novelty and relevance, which novelty is computed by the popularity of papers and relevance is computed by the similarity of papers.