Coupled Collective Matrix Factorization

Collective Matrix Factorization (CMF) makes rating prediction by jointly factorizing multiple matrices in recommender systems (RS), which also provides a unified view of matrix factorization. However, CMF does not directly involve the user attributes and item attributes that represent the intrinsic characteristics of users and items, so it fails to capture the coupling relationships within and between entities, such as users and items, which represent low-level data characteristics and complexities and drive the rating dynamics. In this work, we propose a coupled CMF (CCMF), which not only accommodates entity attributes into rating prediction, but also incorporates the couplings within and between entities into CMF. Therefore, CCMF not only captures the latent variable-based relationships between ratings and specific dimensions at high levels, but also captures the underlying driving forces, i.e., the hierarchical couplings within and between entities representing the low-level data characteristics and complexities. This work also presents a unified framework of CCMF in RS. Experimental results on two real data sets show that our proposed model outperforms the MF-based approaches.