TLSAN: Time-aware long- and short-term attention network for next-item recommendation

Recently, deep neural networks are widely applied in recommender systems for their effectiveness capturing/modeling users' preferences. Especially, the attention mechanism learning enables to incorporate various features an adaptive way. Specifically, as next item recommendation task, we have following three observations: 1) sequential behavior records aggregate at time positions ("time-aggregation"), 2) users personalized taste that is related "time-aggregation" phenomenon ("personalized time-aggregation"), and 3) short-term interests play important role prediction/recommendation. In this paper, propose a new Time-aware Long- Short-term Attention Network (TLSAN) address those observations mentioned above. TLSAN consists of two main components. Firstly, models "personalized time-aggregation" learn user-specific temporal via trainable position embeddings with category-aware correlations long-term behaviors. Secondly, long- feature-wise layers proposed effectively capture preferences accurate recommendation. utilize way, its usage enhances TLSAN's ability dealing sparse interaction data. Extensive experiments conducted on Amazon datasets from different fields (also size), results show outperforms state-of-the-art baselines both capturing performing time-sensitive next-item