Feature Transformation
🎯 Purpose¶
Should You Transform a Feature?¶
✅ Check these conditions:
| Symptom | Suggested Action |
|---|---|
| Strong right-skew in a numeric predictor | Log-transform or Square Root the predictor |
| Predictor has strong non-linear relationship with log-odds | Try binning or adding polynomial features |
| Heavy-tailed distribution (extreme outliers) | Log-transform to compress outliers |
| Large scale differences across features | Standardize (Z-score) or Min-Max scale predictors |
🧠Key Reminders¶
- Logistic regression needs linear relationship between X and log-odds, NOT X and Y directly.
- Target variable (Y) must stay binary (0 or 1) âž” NEVER transform Y.
- Always check distribution plots and log-odds plots if possible.
✅ Transform messy predictors — NOT the binary outcome!
💡 Tip¶
"Before transforming, plot each predictor against the log-odds (or use partial residual plots) to verify non-linearity. This ensures you only transform variables that truly need it, preventing unnecessary complexity and preserving interpretability."