Daily Arxiv

This is a page that curates AI-related papers published worldwide.
All content here is summarized using Google Gemini and operated on a non-profit basis.
Copyright for each paper belongs to the authors and their institutions; please make sure to credit the source when sharing.

A Self-Supervised Mixture-of-Experts Framework for Multi-behavior Recommendation

Created by
  • Haebom

Author

Kyungho Kim, Sunwoo Kim, Geon Lee, Kijung Shin

Outline

This paper addresses the performance improvement of a multi-action recommendation system that leverages various user behaviors (e.g., purchases, clicks, and cart additions) in e-commerce. Existing systems suffer from a significant performance gap between recommended items visited (items users interact with) and unvisited items. To address this, we propose MEMBER, a novel multi-action recommendation system based on a mixture of experts (MIxture-of-Experts), which utilizes expert models specialized for visited and unvisited items, respectively. Each expert model is trained using self-supervised learning, and experimental results demonstrate up to 65.46% performance improvement (based on a hit ratio of 20) compared to existing systems.

Takeaways, Limitations

•
Takeaways:
◦
A novel approach to addressing the performance gap between recommendations for visited and unvisited items.
◦
Combining mixed expert models and self-supervised learning to improve recommendation performance.
◦
Presenting how to effectively utilize various user behavior data.
◦
Significant performance improvements achieved in the Hit Ratio@20 metric.
•
Limitations:
◦
Potential increase in computational costs due to the complexity of the MEMBER model
◦
This is a performance evaluation result for a specific dataset, and generalizability to other datasets needs to be verified.
◦
Additional experimental results on various recommendation indicators are needed.
👍
Made with Slashpage