Unlocking the Predictive Power of Relational Data with Automated Feature Engineering
summary
Uses the H&M Fashion dataset to auto‑generate tabular features for churn classification and item‑sales regression, applying getML FastProp propositionalization over timestamps and join relationships and then training XGBoost/LightGBM. Pipeline handles a 3.8 million‑row population, runs in minutes (example fit ~3m45s), exports the top ~200 features as SQL, and integrates MLflow. Engine is native C++ with memory‑mapped, CPU‑bound execution and benchmarks claim 114–512× faster than TSFresh while matching or beating Relational Deep Learning and skilled humans.
AI-written summary · about
chapters (8)
- 0:00Predictive analytics on relational data
- 1:11Manual feature engineering challenges
- 15:27Automating with getML
- 27:25Benchmarking railbench and H&M
- 32:28Churn prediction pipeline
- 49:00Item sales prediction
- 1:05:27getML versus other approaches
- 1:12:54Q&A
related talks
01h 28mFeature Engineering Made Simple
37mAutomated Machine Learning: Building a Pipeline
01h 39mApplying Rigorous Machine Learning Methods in Business Strategy
42mSolving 3 business problems with ML.NET in 45 minutes
47mBuilding ML practices to address the top four use cases
30mFeature Engineering: An Apprentice’s Guide to the “Dark Art” of Machine Learning
22mPredict the future (or don't) with machine learning in applications
32mUnlock the full predictive power of your multi-table data