Transformer based clustering: Identifying product clusters f...
summary
Claims a production workflow that clusters offers using Sentence-Transformers (mpnet base) fine‑tuned with siamese networks and contrastive/hard‑negative mining, outperforming tf‑idf (0.72) and unfine‑tuned transformers (0.58) and beating arc‑margin (0.82) to reach F0.5 ≈ 0.84 on a 10,000‑product shoe test set (≈17 offers/product). Indexing uses Facebook Faiss, storing 13M vectors in a ~25 GB index; a brute‑force TPU run finishes in ~4 hours (~1.2 ms per vector). The pipeline produces usable clusters but struggles with variant tokens (colors) and noisy titles.
AI-written summary · about
chapters (21)
- 0:00Introduction
- 1:10Who are you
- 1:43About Idio
- 2:29Price Comparison
- 2:56Offers
- 4:15Problem
- 6:48Results
- 16:02Sebastian Wanner
- 27:28Challenges
- 29:25Questions
- 31:18Multilingual knowledge distillation
- 31:52Obtaining ground truth data
- 33:48Manually finding hard negative pairs
- 34:46Wrong labels
- 35:27Treating color words differently
- 36:13Fine tuning data set
- 36:45How often
- 38:11How to recreate experiments
- 39:50Choosing the right metric
- 42:09cosine distance step
- 42:56finetuning
related talks
30mProductionizing an unsupervised machine learning model to...
22mHarnessing image and text data for product matching
49mFrom research to production at ebay
33mNeural Search - Let's talk about quality
25mThe Art of Clustering: The Good, The Bad and The Beautiful
36mFinding needles in haystacks with Deep Neural Networks
39mUnsupervised NLP Classification with Clustering
01h 26mBagging to BERT- A tour of applied NLP