A Self-Attentive Interest Retrieval Recommender

Thanks to the attention mechanism, self-attention networks (SANs) have been widely used in sequential recommendation. However, most existing SANs approaches still follow an old fashion generating one single embedding as final representation, which constraints model's capacity. To enrich this kind of representation, sequential recommender uses metadata such as item category to capture user's multi-interests. But this method will not reach its expectation due to item's long-tail property. This property will result a large constant of category cannot be effectively activated by the lack of interaction records. Another drawback is that may also lead to over-parameterization caused by the massive categories. Particularly, we propose a Self-Attentive Interest Retrieval network (SAIR) to explore a context-aware representation from user's behaviors while not fall into over-parameterization. SAIR works in a typical SANs manner, encode the behavior sequence using self-attention, and we propose an interest retrieval module to project the sequences to an interest relevance distribution adaptively. And we leverage an interest-to-interest interaction to generate several context-aware interests embeddings. Then we fuse multi-interest embeddings as final output. Extensive experiments are carried out on three real-world datasets, the results demonstrate that SAIR outperforms other SANs methods and other state-of-the-art algorithms in multiple evaluation metrics.