A study in classification
summary
Builds an ML pipeline that assigns Harmonized System (HS) classification codes using scraped international shipping records, third‑party aggregators and feature engineering on name description and material fields. Tries sklearn models, fastText and deep nets, uses prodigy for rapid labelled examples and a human-in-the-loop workflow. Initial models hit about 40% to 80% accuracy during iteration, with real human repeatability near 70%, so the system succeeds as an augmenting classifier that speeds decisions and raises consistency rather than fully replacing human review.
AI-written summary · about
chapters (14)
- 0:00Introduction
- 0:32About Semantic3
- 1:23Use Case
- 3:57HS Codes
- 5:05Problem Statement
- 8:25Collecting Data
- 11:16Gradient Descent
- 12:25Validation
- 13:24Assumptions
- 15:57Output structure
- 17:24Predictions imitate data
- 19:41Cost
- 20:39Data
- 23:50Evolution
related talks
38mE-Commerce Product Categorization at Scale
44mContent Classification: NLP vs. AI in the Age of Data Vectorization
18mData Classification at Scale: Taming the Hydra
37mAutomated Machine Learning: Building a Pipeline
01h 25mUnlocking the Predictive Power of Relational Data with Automated Feature Engineering
15mMachine learning: Necessity vs obligation
01h 39mApplying Rigorous Machine Learning Methods in Business Strategy
27mUsing Machine Learning to solve a classification problem with scikit-learn