A Bottom-Up Approach to Building a Product Taxonomy
summary
Mirakl describes a bottom-up method to build product taxonomies from millions of real product records instead of imposing a top-down structure. They combine product embeddings and large language models to cluster products into domain-specific categories and aggregate observed product attributes to identify the most relevant attributes per category, and they explain the system architecture and technical trade-offs that make this approach suited to many industries.
AI-written summary · about
chapters (12)
- 0:00Product taxonomy intro
- 3:39Ideal taxonomy criteria
- 6:06Top-down vs bottom-up
- 8:56Product data collection
- 11:04Automatic category creation
- 15:19Attribute extraction and deduplication
- 17:26Taxonomy evaluation results
- 21:17Results vs standards analysis
- 26:07Mapping pivot role
- 28:32Project limits and learnings
- 32:42Bottom-up advantages
- 34:49Conclusion & Q&A
related talks
32mLessons Learnt Building Domain Specific NLP Pipelines
18mData Classification at Scale: Taming the Hydra
42mUsing Interconnected ML Models to Tackle Retail Challenges
52mA Million Products, One Experience: Using Collaborative and Multimodal Embeddings
22mBefore the Model: How Machine Learning Products Start...
46mWhy data-driven methods will shape the future (...)
37mBuilding a Scientific Taxonomy at Scale
38mE-Commerce Product Categorization at Scale