A Generic Framework for Learning Explicit and Implicit User-Item Couplings in Recommendation

The nature of recommendation is Non-IID, which has potential in improving recommendation quality and addressing issues such as sparsity and cold start. However, existing many state-of-the-art methods assume users and items are independent and same distributed while ignoring complex coupling relationships within and between users and items, resulting in limited performance improvement. To solve this issue, this paper proposes a novel neural user-item coupling learning model, short for CoupledCF, based on non-IID learning for collaborative filtering. CoupledCF joint learns explicit coupling with CNN and implicit coupling with deepCF within/between users and items accompanying user/item side information for recommendation tasks. User/item side information contains of attribute-based and feature-based. For different user/item side information, we use different embedding methods to learn embedding representation. We conduct comparative experiments on (1) two datasets from MovieLens1M and Tafeng with attribute-based user/item information for Top-K recommendation. (2) two datasets from MovieLens1M and BookCrossing with attribute-based user/item information for rating prediction. (3) two datasets from Amazon Movies and TV (AMT) and Yelp for feature-based user/item information for Top-K item recommendation and rating prediction tasks. Empirical results on five available real-world large datasets prove our proposed CoupledCF model is able to obtain better recommendation accuracy compared with several mainstream approaches for recommendation: BMF, neural matrix factorization, Google's Wide&Deep network, DeepFM, convMF, and A³NCF model.