Collaborative filtering enhanced by user free-text reviews topic modelling

Abstract

User based collaborative filtering (UCF) and item based collaborative filtering (ICF) are two quite successful approaches applied in recommender system, both of which try to predict the rating value of items for the target user based on the items previously rated by other users. However, existing UCF and ICF base on only a rating and neglect the fact that users' same rating to an item may be based on different item features. Another deficiency of UCF and ICF is that both approaches are limited due to the data sparsity problem. In this paper, a hybrid collaborative filtering recommender system that enhances traditional CF (UCF, ICF) by user's free-text review information is presented. Latent Dirichlet Allocation (LDA) is used in our study to infer the topic probability distribution of the users' reviews and then we propose two aggregation methods to uncover user-topic-interest profile and item-topic-features profile, which will be used to measure similarity between users and between items. Finally we try to combine this new proposed similarity measurement with that of traditional CF. Experiments we conducted show that our algorithms perform better than UCF, ICF and the popular recommendation algorithm Slope One on datasets with different sparsity.