Harnessing image and text data for product matching
summary
Claims a working end to end product‑matching pipeline that scales to ~40 million products and roughly 200 million images, using off‑the‑shelf vision transformers from Hugging Face to create embeddings, Faiss/Vector DB style approximate nearest‑neighbor search, then a sentence‑pair classification model on titles to resolve bundles and spec differences. Built outputs include an internal real‑time visual search tool and automated duplicate detection that already merges listings and informs competitive pricing; experiments and A/B tests run to validate conversion impact.
AI-written summary · about
chapters (9)
- 0:00Product matching challenges
- 2:25Scalable matching pipeline
- 6:39Embeddings for semantic similarity
- 9:50Vector search with ANN
- 12:35Image-text for precise matching
- 15:47Duplicate product detection
- 16:50Image quality with embeddings
- 18:57Visual product search tool
- 21:12Summary
related talks
41mSearching for Products by Photo
33mMachine Learning in Visual Search
22mMachine Learning: The Engine Behind Modern E-commerce Search
36mFinding needles in haystacks with Deep Neural Networks
29mDeveloping Visual AI Solutions for Online Marketplace
30mEvolving Search in Online Marketplaces: Practical Applications of AI
39mAI-Powered Search in Modern E-Commerce
37mSimilarity Search for Product Matching @ Semantics3