Feature engineering
Where most of the accuracy actually comes from on tabular problems, and where most production failures are born.
| # | File | The question it answers |
|---|---|---|
| 01 | Encoding categorical features | how to encode a 40,000-value column without one-hot |
| 02 | Scaling and numeric transforms | which models need scaling, skew, outliers, missing values |
| 03 | Feature selection | which features to keep — including operational cost, not just statistics |
| 04 | Data leakage | the most expensive bug in ML |
| 05 | Imbalanced data | 99:1 classes without reaching for SMOTE first |
| 06 | Feature stores and point-in-time correctness | training/serving skew and point-in-time correctness |
Read 04 first. Leakage doesn’t crash, doesn’t warn, and makes your metrics better — which is exactly why it reaches production. The reflex to build is: unexpectedly good results are leakage until proven otherwise.