# Attention over Self-attention:Intention-aware Re-ranking with Dynamic Transformer Encoders for Recommendation

**Authors:** Zhuoyi Lin, Sheng Zang, Rundong Wang, Zhu Sun, J. Senthilnath, Chi Xu, Chee-Keong Kwoh  
**arXiv:** [2201.05333](https://arxiv.org/abs/2201.05333)  
**Date:** 2022-01-14  
**Summary page:** https://www.summarizepaper.com/en/paper/2201.05333v3/

## Key points

- Re-ranking models enhance item recommendations by refining lists generated by global ranking models
- Existing re-ranking solutions often rely solely on implicit feedback and a shared prediction model, overlooking intricate inter-item relationships within diverse user intentions
- The RAISE model introduces a groundbreaking approach to provide user-specific predictions tailored to individual users based on their unique intentions
- RAISE extracts latent user intentions from text reviews using an intention discovering module (IDM) and captures them for each user-item pair with a co-attention network
- RAISE incorporates a dynamic transformer encoder (DTE) to capture user-specific inter-item relationships among potential item candidates by integrating learned latent user intentions obtained through IDM
- Empirical evaluations demonstrate superior performance of RAISE compared to existing methods, achieving significant relative improvements in key evaluation metrics such as Precision@5, MAP@5, and NDCG@5 up to 13.95%, 9.60%, and 13.03% respectively
- Authored by Zhuoyi Lin, Sheng Zang, Rundong Wang, Zhu Sun, J. Senthilnath, Chi Xu, and Chee-Keong Kwoh and accepted by IEEE TKDE

## Summary

In the field of recommendation systems, re-ranking models have been proven to enhance the quality of item recommendations by refining lists generated by global ranking models. However, existing re-ranking solutions often rely solely on implicit feedback and a shared prediction model, overlooking the intricate inter-item relationships that exist within diverse user intentions. To address this limitation, this paper introduces a groundbreaking approach known as the . The RAISE model is designed to provide user-specific predictions tailored to individual users based on their unique intentions. The process begins with the extraction of latent user intentions from text reviews using an intention discovering module (IDM). By leveraging a co-attention network to differentiate the importance of review information, the latent user intention is explicitly captured for each user-item pair. Furthermore, RAISE incorporates a dynamic transformer encoder (DTE) to capture user-specific inter-item relationships among potential item candidates. This is achieved by seamlessly integrating the learned latent user intentions obtained through IDM. As a result, RAISE not only delivers more personalized recommendations but also offers corresponding explanations by building upon existing recommendation engines. Empirical evaluations conducted on four public datasets demonstrate the superior performance of RAISE compared to existing methods. The proposed model achieves significant relative improvements in key evaluation metrics such as Precision@5, MAP@5, and NDCG@5, with enhancements reaching up to 13.95%, 9.60%, and 13.03% respectively. Authored by Zhuoyi Lin, Sheng Zang, Rundong Wang, Zhu Sun, J. Senthilnath, Chi Xu, and Chee-Keong Kwoh and accepted by IEEE TKDE, this research showcases how attention over self-attention techniques can revolutionize re-ranking strategies in recommendation systems for enhanced user experiences and satisfaction levels.