You Like What You Hear: Improving Movie Recommender Systems Using Textual Information in Movie Transcripts
Rong Jian Zheng, Alex Tuzhilin · Rare & Special e-Zone (The Hong Kong University of Science and Technology) · 2016
We study how to use the deep content information contained in a movies’ transcript to improve the predictive performance of the movie recommender system. We propose to use Latent Dirichlet Allocation (LDA), an unsupervised topic generation technique, to extract the thematic content of movies at the topic level. We argue that the extracted topic distribution for a movie is related to a consumer’s personal taste and therefore influence his or her ratings about the movie. We then use movies’ topic distributions to quantify the similarities between any pair of movies. We compare the performance of the proposed topic-based collaborative filtering system with several benchmark systems including item-based, user-based collaborative filtering and matrix factorization methods. We find that the proposed approach is more effective than the benchmark systems when the rating data set is sparse or when the proportion of new movies is. Our approach is therefore a solution to alleviate the notorious cold-start problem in the movie recommender system.