부산대학교 도서관

Accurate representation of user behavior characteristics is one of the key factors to improve the accuracy of collaborative filtering recommendation. However, only using sparse rating data to represent user behavior greatly deteriorates the performance of the recommender system. Extracting more effective features of auxiliary textual information, such as reviews or abstracts, not only can effectively alleviate data sparsity, but also can better improve prediction accuracy. This paper proposes a novel model, namely TCLPMF, which integrates textual Convolutional Neural Network and Latent Dirichlet Allocation to capture milti-level features into a probabilistic matrix factorization model. Extensive experiments show that compared with the traditional recommendation model, such as PMF and ConMF, the mean absolute error of the proposed model are reduced on Movielens dataset.