HAN: Hierarchical Attention Network for Learning Latent Context-Aware User Preferences With Attribute Awareness

Sequential recommendation has become a key role in many online services as user activities grow. With recent advancements in the prediction of users' next interests through joint modeling of long-term and short-term preferences, recommendation performance has significantly improved. However, existing work largely overlooks the evolving nature of users' preferences, directly integrating interactions between item-item or user-item features, with no proper understanding of what motivates the user in the selection of a sequence. Users usually make their decisions on a two-step basis: initiating an action on a specific intent and then selecting an item that satisfies the intent. Besides, the existing models fail to properly capture the hierarchical relationships between multiple attributes of items and the temporal dependencies of the items within sequences. In this paper, we introduce a novel architecture for a sequential recommendation that unifies users' long- and short-term preferences and models the underlying user intentions. Our model employs a Hierarchical Attention Network (HAN), which can grasp fine-grained relationships between item attributes in sessions and incorporates both temporal and semantic dependencies among items. By explicitly modeling user intentions and their evolving preferences, the HAN illustrates better how users act in behavior. On two real datasets, we show that the proposed model HAN outperforms all other current state-of-the-art methods.