Abstract: A dual-aspect item attention network (DIAN) for item-based recommendation is proposed, which jointly takes into account the aspects of importance of historical items in a user profile to the target items and the underlying relations among these items. DIAN consists of two main modules, a neural attentive model for item similarity between historical and target items (NAIS), and a dual normalization self-attention item similarity model for item similarity underlying historical items (SAIS). On one hand, the neural attentive model is introduced to distinguish the different contribution of the historical items in a user profile to the perdition on the target item. On the other hand, a self-attention network is proposed to infer the item-item relationship from users’ historical interactions, which is able to estimate the relative weights of each item in user interaction trajectories, in order to learn better representations for users’ interests. Furthermore, a self-attention network is proposed using a dual normalization mechanism, consisting of a layer focusing on extracting users’ representation from historical items, and a layer making it unaffected by the number of users’ historical items. Extensive experiments conducted on two public benchmarks demonstrate the proposed method outperforms the state-of-the-art recommendation models.

Key words: attention networks, collaborative filtering, self-attention model, dual normalization, recommendation