MemRerank: Preference Memory for Personalized Product Reranking
Summary: arXiv:2603.29247v1 Announce Type: cross
Abstract: LLM-based shopping agents increasingly rely on long purchase histories and multi-turn interactions for personalization, yet naively appending raw history to prompts is often ineffective due to noise, length, and relevance mismatch. We propose MemRerank, a preference memory framework that distills user purchase history into concise, query-independent signals for personalized product reranking.
Introduction
In the realm of e-commerce, the personalization of shopping experiences has become increasingly important. As machine learning models, particularly large language models (LLMs), are deployed in shopping agents, they are designed to utilize extensive purchase histories along with multi-turn interactions to enhance user experience. However, a challenge arises when attempting to incorporate these long histories into the models, as simply appending raw data can lead to inefficiencies and inaccuracies.
The Challenge of Personalization
Traditional methods of utilizing user purchase history may lead to:
- Noise: Irrelevant or outdated information can clutter the decision-making process.
- Length: Too much data can overwhelm the model and hinder its performance.
- Relevance mismatch: Important signals may be lost among less relevant information.
Introducing MemRerank
To address these challenges, we introduce MemRerank, a sophisticated framework that transforms user purchase histories into compact, query-independent signals. This approach aims to facilitate more efficient and effective personalized product reranking. The key features of MemRerank include:
- Preference Memory: Distillation of user preferences into manageable data representations.
- End-to-End Benchmarking: Construction of a robust evaluation framework centered around a 1-in-5 selection task.
- Reinforcement Learning: Training of the memory extractor to optimize downstream reranking performance.
Evaluation Framework
Our evaluation framework is designed to measure two primary aspects:
- Memory Quality: Assessing how well the preference memory captures user intent.
- Reranking Utility: Evaluating the effectiveness of the reranking process in presenting personalized options.
Experimental Results
In our experiments, we compared MemRerank against various baselines, including no-memory, raw-history, and off-the-shelf memory approaches. The results were compelling, demonstrating that MemRerank consistently outperformed these alternatives:
- Achieving up to +10.61 absolute points in 1-in-5 accuracy.
- Proving that explicit preference memory significantly enhances personalization in e-commerce systems.
Conclusion
MemRerank represents a significant advancement in the field of personalized e-commerce. By effectively distilling user purchase history into actionable insights, it not only improves the accuracy of product recommendations but also enhances the overall shopping experience. As LLM-based agents continue to evolve, MemRerank stands as a practical and effective solution for building personalized, agentic shopping systems.