Reproduce “MRSE: An Efficient Multi-modality Retrieval System for Large Scale E-commerce” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Hao Jiang, Haoxiang Zhang, Qingshan Hou, Chaofeng Chen, Weisi Lin, Jingchang Zhang, Annan Wang
An efficient multi-modality retrieval system (MRSE) was developed to improve item recall quality in large-scale e-commerce by integrating text, image, and user preferences. Leveraging a Lightweight Mixture-of-Experts architecture and a hybrid loss function, the system achieved up to a 34.7% increase in offline Recall@K and a 3.2% increase in Gross Merchandise Volume during online A/B testing on Shopee.
Key insights
- Explicitly modeling individual user preferences across modalities using a Lightweight Mixture-of-Experts (LMoE) architecture significantly enhances retrieval relevance by adapting to diverse user needs.
- A hybrid loss function that strategically combines easy (mix-sampled) and hard negative examples is crucial for achieving both rapid convergence and improved model robustness in large-scale multi-modality ERS.
- A comprehensive multi-modal representation, incorporating image, category, historical, and behavioral features in addition to textual data, collectively and substantially improves item recall and relevance in e-commerce search.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.