E-Commerce Product Categorization at Scale
summary
Honey builds an ensemble that combines a text CNN (title + merchant tokens), an image transfer model (Inception backbone), Airflow orchestration and model training on Google AI Platform, plus regex and store rules fed to a voting master. Coverage includes ~27,000 manual mappings (56% product coverage) and ~20 million training examples; a 77k holdout across 3,100 categories shows high precision at coarse levels but only ~50% recall, so uncertain items go to crowdsourcing; image training is paused due to TFRecord / quota issues.
AI-written summary · about
chapters (11)
- 0:00Introduction to Honey
- 3:18Need for product categorization
- 6:50Manual mapping approach
- 8:57Human-machine hybrid solution
- 11:24Model ensemble architecture
- 14:09Automated training pipeline
- 17:37Store and regex rules
- 19:04Voting master and performance
- 21:11Crowdsourcing validation
- 22:57Key takeaways and future plans
- 27:39Q&A session
related talks
22mHarnessing image and text data for product matching
30mA study in classification
34mComputer Vision and NLP for Multi-Task Fashion Attrb. Modeling
22mMachine Learning: The Engine Behind Modern E-commerce Search
42mUsing Interconnected ML Models to Tackle Retail Challenges
36mFinding needles in haystacks with Deep Neural Networks
28mOptimizing Text Classification in Construction Data
29mHow GenAI is transforming e-commerce catalogues