Cross-Domain Collaborative Filtering with Review Text

Most existing cross-domain recommendation algorithms focus on modeling ratings, while ignoring review texts. The review text, however, contains rich information, which can be utilized to alleviate
data sparsity limitations, and interpret transfer patterns. In this paper, we investigate how to utilize the review text to improve cross-domain collaborative filtering models. The challenge lies in the existence of non-linear properties in some transfer patterns. Given this, we extend previous transfer learning models in collaborative filtering, from linear mapping functions to non-linear ones, and propose a cross-domain recommendation framework with the review text incorporated. Experimental verifications have demonstrated, for new users with sparse feedback, utilizing the review text obtains 10% improvement in the AUC metric, and the nonlinear method outperforms the linear ones by 4%.