Robust AI Search Ranking for Radical C2C Marketplace Growth
summary
Mercari deploys a decoupled post-retrieval ML reranker (ElasticSearch retrieval + MLP learning-to-rank) serving tens of thousands QPS with p99 latency in the low hundreds ms, trained on implicit labels from 23M MAU and 3B listings. Key moves: drop noisy clicks, filter to high-signal purchase labels and train on four weeks + one week validation, use intervention harvesting to estimate propensities from A/B variants, and ship models via TensorFlow Ranking and Weights & Biases; purchase NDCG climbs and yields consistent AB wins.
AI-written summary · about
chapters (9)
- 0:00Introduction and overview
- 1:21Mercari platform context
- 2:59Search and reranking system
- 7:15Data construction and eval
- 13:15Advanced data strategies
- 18:18Model building architecture
- 20:34Data debiasing techniques
- 30:29Takeaways and impact
- 33:22Audience Q&A
related talks
41mBuilding MLOps Infrastructure at Japan's Largest C2C E-Commerce Site
20mAuto Content Moderation in C2C e-Commerce
39mRanking Listings with Machine Learning and ElasticSearch
39mAI-Powered Search in Modern E-Commerce
22mMachine Learning: The Engine Behind Modern E-commerce Search
31mEverything in its Right Place: Optimising Ranking in Online Grocery
30mEvolving Search in Online Marketplaces: Practical Applications of AI
33mAI-Powered Search in Modern E-Commerce