Machine learning: Necessity vs obligation
summary
Retail data cleaning compares a simple rule table and rule-based tree against a language transformer model: pack-count prediction moves from 79% to 85% accuracy (+6 points in their test) by automating pattern discovery, while brand ID across ~86,000 labels sees rule logic and lookup/string-matching hit ~77% but the transformer falls to ~55% because rare brands lack samples. Result: automated pack/brand workflows cover roughly 20–49% of incoming items, cut human labeling and speed retailer onboarding, and avoid becoming one of the 87% failed ML projects.
AI-written summary · about
chapters (6)
- 0:00Start
- 2:35Spin's Mission and Use Cases
- 3:21Use Cases for Ml
- 9:46Use Case of Brand Identification
- 10:04Rule-Based Tree Model
- 11:47Takeaways from the Two Use Cases
related talks
36mMachine Learning Sucks
28mHindsight bias: A machine learning story
16mGen AI For Retail Data Management
38mPlaying by the Rules-Based-Systems
37mMachine Learning, The Right Way
13mHarnessing Machine Learning to Combat E-commerce Fraud
27mScalable Data Pipelines - an ML Approach
25mAutomating Operations with ML