Incorporate Topic Model Into Collaborative Filtering

With the development of information technology and Internet, we are coming into a highly informationization era. Traditional search engines cannot satisfy the requirements of users and e-commerces simultaneously. To this end, recommender systems came to being.However, most existing recommender systems only take rating scores for granted and discard the wealth of information in content related to users or items, such as items’ content, users’ reviews etc. There are abundant information about users’ preferences embedded in these content, which is very significant for improving recommender systems. In the field of text modeling, topic model(e.g., LDA) is the most successful statistical learning method. Therefore, we focus on incorporating topic model into collaborative filtering based recommender systems.First, in order to exploit user preferences’ information embedded in both ratings and reviews exhaustively, we proposed a Bayesian recommender model, named User Rating and Review Profiling(URRP), which links a traditional collaborative filtering technique with a topic model seamlessly. With review text information involved, latent user rating attitudes are interpretable and “cold-start” problem can be alleviated. Experimental results on 25 realworld datasets demonstrate the superiority of our URRP model over state-of-the-art methods.Futhermore, individuals’ interests and concerning topics are generally changing over time,with strong impact on their behaviors in social media. Accordingly, designing an intelligent recommender system which can adapt with the temporal characters of both factors becomes a significant research task. To this end, we proposed a novel dynamic recommender model based on collective factorization, named Temporal and Topic-Aware Recommender Model(TTARM). Experimental results on two real life datasets from CiteULike and MovieLens demonstrate the effectiveness of TTARM.