# Do Not Wait: Learning Re-Ranking Model Without User Feedback At Serving Time in E-Commerce

**Authors:** Yuan Wang, Zhiyu Li, Changshuo Zhang, Sirui Chen, Xiao Zhang, Jun Xu, Quan Lin  
**arXiv:** [2406.14004](https://arxiv.org/abs/2406.14004)  
**Date:** 2024-06-20  
**Summary page:** https://www.summarizepaper.com/en/paper/2406.14004v1/

## Key points

- Recommender systems are essential in e-commerce for providing personalized product recommendations.
- Re-ranking models play a crucial role in shaping final recommendation lists by utilizing inter-item influence.
- Online learning methods update models with the latest data to capture shifts in data distributions.
- A challenge arises from delays in real-time user feedback, impacting model enhancement and responsiveness.
- The LAST (Learning At Serving Time) approach revolutionizes re-ranking modeling by eliminating the need for real-time user feedback.
- LAST dynamically applies request-specific model modifications on-the-fly to generate tailored recommendations without error propagation.
- LAST surpasses state-of-the-art re-ranking models in performance and efficiency through comprehensive experiments both offline and online.

## Summary

In the realm of e-commerce, recommender systems have become indispensable tools for guiding users towards personalized product recommendations. Among these systems, re-ranking models have emerged as key players in shaping the final recommendation lists by harnessing inter-item influence. To keep pace with the ever-evolving data landscape of e-commerce, online learning methods continuously update deployed models with the latest samples to capture shifts in underlying data distributions. However, a significant challenge arises from the reliance on real-time user feedback, which can be delayed by hours or even days, particularly in cases like item purchases. This delay often results in a lag in model enhancement and responsiveness. Addressing this challenge head-on, a groundbreaking approach known as LAST (Learning At Serving Time) has been introduced to revolutionize re-ranking modeling without depending on user feedback at serving time. By leveraging a surrogate model to provide instructional signals for model improvement, LAST eliminates the need for real-time user feedback. When an online request is received, LAST dynamically applies request-specific and transient model modifications on-the-fly before generating tailored recommendation results for that specific request. Importantly, these modifications are discarded after each request to prevent error propagation and stabilize the online learning process. As a complementary addition to feedback-based online learning methods, LAST seamlessly integrates into existing systems to enhance adaptability and responsiveness in recommendation experiences. Through comprehensive experiments conducted both offline and online, it has been demonstrated that LAST surpasses state-of-the-art re-ranking models in terms of performance and efficiency. With its innovative approach to enhancing re-ranking models without user feedback delays, LAST stands out as a game-changer in optimizing recommendation systems within the dynamic landscape of e-commerce.