Reproduce “AIGQ: An End-to-End Hybrid Generative Architecture for E-commerce Query Recommendation” 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
Jingcao Xu, Jianyun Zou, Renkai Yang, Zili Geng, Qiang Liu, Haihong Tang
AIGQ, a hybrid generative architecture developed by Alibaba's Taobao & Tmall Group, addresses personalized pre-search query recommendation for e-commerce by integrating interest-aware list training and continuous policy optimization. The framework achieved online increases of over 10% in attributed orders and Gross Merchandise Volume (GMV) on the Taobao platform.
Key insights
- Explicit list-aware supervised fine-tuning and fine-grained reinforcement learning, enhanced with dual-level rewards, are complementary and critical for optimizing ordered list generation in LLM-based recommenders.
- Reasoning augmentation (AIGQ-Think) improves generation accuracy and diversity by distilling complex user intent into coherent interest points before generating queries.
- Continuous policy optimization through daily reinforcement learning updates, integrating a CTR reward from production models, effectively aligns the LLM's generation policy with real-time user preferences and mitigates distribution shift.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.