📋 Linear Feature Transformation Trigger Card
🎯 Purpose¶
Use this card to determine when to apply feature transformations (e.g. log, sqrt, power, binning) before running a linear regression model. Helps identify skewed or nonlinear predictors that may violate model assumptions or degrade performance.
🔍 1. When to Transform Features¶
| Condition | Suggestion |
|---|---|
| Skewness > ±1 | Use log or sqrt transform |
| Heteroskedasticity observed | Log/sqrt to stabilize variance |
| Long-tailed distribution | Consider log or Box-Cox/Yeo-Johnson |
| X vs Y non-linear | Try polynomial or log(X) variants |
| Zero-inflated or boundary effects | Apply binning or thresholding |
from scipy.stats import skew
skew(df['income']) # > 1 → log candidate
🧪 2. Visual Triggers¶
| Visual | Transformation Cue |
|---|---|
| Histogram with long right tail | Log or sqrt |
| Residual plot shows fan shape | Variance-stabilizing transform |
| X vs Y: curvature in scatter | Poly features or log(X) |
| Extreme outliers (Z > 3) | Consider robust scale or winsorization |
⚖️ 3. Transformation Options¶
| Method | Use When... |
|---|---|
np.log1p(X) |
Right-skewed + zero values present |
np.sqrt(X) |
Moderate skew, non-negative |
power_transform() |
General normalizer (Box-Cox or YJ) |
StandardScaler |
When model is sensitive to scale (esp. Ridge/Lasso) |
✔️ Scale after transforming, especially if using regularized models
✅ Transformation Checklist¶
- [ ] Skew > ±1 or kurtosis > 3 reviewed
- [ ] Visual cues support transformation
- [ ] Chosen method appropriate for data range
- [ ] Post-transform distribution checked
- [ ] Feature scaling applied after transformation
💡 Tip¶
“Linear models love smooth, symmetric input. If your predictors are shouting, transformation helps them whisper.”