MH-NAIS: A Multi-Head Attention Extension of the Neural Attentive Item Similarity Model

The Neural Attentive Item Similarity (NAIS) model shows strong performance in identifying item-item relationships using attention mechanisms [1]. However, its single-head attention mechanism restricts the model’s ability to capture the diversity of user preferences and complex interaction patterns. To address this limitation, we propose MHNAIS, an enhanced version of NAIS that integrates a multi-head attention mechanism to improve model expressiveness and robustness. By enabling the model to attend to multiple representation subspaces in parallel, MH-NAIS more effectively captures nuanced semantic signals from users’ historical interactions. Experimental results on the MovieLens-1M dataset show that MH-NAIS consistently outperforms the baseline NAIS model, achieving up to $3.0 \%$ higher Hit Ratio (HR@10) and 4.2% higher Normalized Discounted Cumulative Gain (NDCG@10) in top-N recommendation tasks.