Reproduce “Embedding-based Product Retrieval in Taobao Search” 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
Sen Li, Fuyu Lv, Taiwei Jin, Guli Lin, Keping Yang, Xiaoyi Zeng, Xiao-Ming Wu, Qianli Ma
Alibaba Group's Multi-Grained Deep Semantic Product Retrieval (MGDSPR) model enhances e-commerce search by generating more relevant product embeddings and aligning training objectives with inference. This led to a +0.77% increase in Gross Merchandise Volume (GMV) and a +22.53% increase in products proceeding to pre-ranking on Taobao, addressing challenges of low retrieval relevance and training-inference discrepancy.
Key insights
- Aligning training objectives (using softmax cross-entropy loss) with online inference (global top-K retrieval) significantly improves model performance and accelerates convergence in large-scale embedding-based retrieval systems.
- Smoothing noisy implicit feedback with a temperature parameter and generating relevance-improving hard negative samples are effective and practical methods to enhance the precision and relevance of learned product embeddings.
- A hybrid approach that combines powerful semantic matching from deep learning with a post-retrieval exact match relevance control module is crucial for achieving both high relevance and broad coverage in large-scale e-commerce search.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.