Deep Item-based Collaborative Filtering for Top-N Recommendation

Item-based Collaborative Filtering (ICF) has been widely adopted in recommender systems due to its strength in modeling user interests and ease of online personalization. Existing learning-based ICF methods largely consider only linear and shallow relationships between items, which are insufficient to capture users’ complex decision-making. This work proposes DeepICF, a more expressive ICF solution that models nonlinear and higher-order item relationships. Beyond pairwise (second-order) item interactions, it considers interactions among all historical item pairs via neural networks, enabling differentiation of which itemsets in a user’s profile are more influential for predicting a target item. Experiments on MovieLens and Pinterest verify the positive effect of modeling higher-order interactions with nonlinear neural networks. Furthermore, integrating an attention mechanism to refine second-order interaction modeling (DeepICF+a) yields additional performance gains.