부산대학교 도서관

Collaborative filtering (CF) faces two challenges for recommendations: data sparsity and cold-start issue. One solution is to incorporate the side information and the other is to utilize relevant knowledge. In this paper, a cross-domain deep collaborative filtering (CDDCF) model is proposed by considering both, which combines matrix tri-factorization and deep structure in both source and target domains. Deep structure takes as input the side information to learn latent representation. Matrix tri-factorization generates private latent factors connecting deep structure and common latent factors bridging relevant domains. Experiments on real datasets demonstrate its effectiveness.