Feature-Level Attentive ICF for Recommendation

Item-based collaborative filtering (ICF) enjoys the advantages of high recommendation accuracy and ease in online penalization thus is favored by industrial recommender systems. ICF recommends items to a target user based on their similarities previously interacted user. Great progresses have been achieved for recent years applying advanced machine learning techniques (e.g., deep neural networks) learn item similarity from data. The early methods simply treat all historical equally recently proposed attempt distinguish different importance when recommending item. Despite progress, we argue that those models neglect diverse intents users adopting watching movie because director, leading actors, or visual effects). As result, they fail estimate finer-grained level predict user’s preference an item, resulting sub-optimal recommendation. In this work, propose general feature-level attention method models. key our factors computing prediction. To demonstrate effectiveness method, design light network integrate both item-level It model-agnostic easy-to-implement. We apply it two baseline evaluate its six public datasets. Extensive experiments show enhanced consistently outperform counterparts, demonstrating potential differentiating