CLSAN:category-aware Long-and Short-termAttention Network for SequentialRecommendation

Abstract Sequence recommendation has become a hot research topic due to its practical applications and high precision in obtaining sequence information. However, existing methods often lack consideration for both the global stability and local volatility of user preferences, resulting in decreased learning of the user’s current preferences. Additionally, the accuracy of recommendations is affected by the high sparsity of user check-in data. To address these challenges, we propose a categoryaware user long- and short-term preference recommendation model, CLSAN. Firstly, item category information embedding is introduced in the embedding layer to capture the relevance of item information and effectively alleviate data sparsity. Secondly, to capture the time dependency and user preferences between sessions, the combination of item category embedding and learnable position embedding is fed into the multi-head self-attention module and long- and short-term feature attention layers. Then, the feature attention layer combines this information with user embedding information to generate the user’s final profile. Finally, the generated user preference information is fed into the prediction layer for the recommendation. We evaluate the proposed model on ten real datasets publicly available on Amazon, and the experimental results show that the CLSAN model performs significantly better than existing methods.