Cross Attention Integration: Enhancing Session-based Recommendations with Item-Category Interactions

Eunjo Jang, Yeonju Song, Ki Yong Lee

Abstract: Session-based Recommendation (SR) presents significant challenges, mainly when user interaction data is limited. Traditional SR methods often focus on item IDs, neglecting the intricate relationships between items and their respective categories. Recent research has attempted to address this limitation by incorporating category information alongside item IDs using self-attention mechanisms. However, these approaches have their own set of limitations. Firstly, they risk losing essential information while integrating item and category embeddings into a single vector before the attention layer. Secondly, when utilizing separate attention layers for items and categories, these methods may excel in capturing intra-item or intra-category relationships. However, they might struggle to represent the direct interactions between items and categories accurately. Lastly, these methods often neglect the category's value when incorporating category information into the attention layer, relying solely on the category's query or key. To overcome these limitations, we introduce CAT-SR, a novel cross-attention approach for Session-based Recommendation. CAT-SR integrates item and category information within the attention layer through a unique cross-attention mechanism. This empowers CAT-SR to effectively capture the intricate interactions between items and categories, a crucial factor in delivering precise and contextually relevant Session-based Recommendations. Our empirical evaluations on two real-world datasets demonstrate that CAT-SR outperforms state-of-the-art Session-based Recommendation models.