Fusion topic model transfer learning for cross-domain recommendation

摘要

It remains a challenge to deal with the diversity of the cross-domain feature space when using transfer learning in the recommendation system. To solve the difficulty, we propose a fusion topic model to extract the latent topic in the cross-domain. There are two layers in the proposed model. Firstly, the model simulates the user-item relationship in every subdomain with an author-topic model separately, and extracts the subdomain level topics. In addition, the model extracts the full-domain level topics in the whole domain using subdomain level topics as words in the author-topic model. By using Gibbs sampling, the method can extract two different levels of topics. The experiment on the public dataset shows our method has good performance. The results of the experiment indicate that extracting multilevel topics can help to discover the correlation in the MovieLens dataset and the Book-Crossing dataset, and to extract the cross-domain feature space.