Feature Transformation (Logistic)
🎯 Purpose
Use this card to decide when to apply transformations when working with logistic regression, Poisson, or negative binomial models. Focuses on skew handling, scale sensitivity, and model assumptions.
🔍 1. Trigger Logic by Feature Type
🔹 Numeric (Continuous)
| Condition |
Transformation |
|
|
| Right-skewed distribution ( |
skew |
> 1) |
✅ Log or Sqrt Transform |
| Wide magnitude spread |
✅ StandardScaler or MinMaxScaler (esp. for regularized logistic) |
|
|
Count-based exposure (e.g., minutes, visits) |
✅ Log or Offset transform (Poisson/Negative Binomial) |
|
|
🔹 Categorical
| Condition |
Transformation |
| Nominal (unordered) |
✅ One-hot Encoding |
| Ordinal (known rank) |
✅ Ordinal Encoding |
| High-cardinality |
🔁 Grouping, embedding, or collapse (manual) |
🧪 2. Model-Specific Cues
| Model |
Trigger |
Action |
| Logistic Regression |
Skewed numeric, regularization |
Scale or log |
| Poisson Regression |
Log-linear fit assumption |
Log transform predictors + offset(optional) |
| Negative Binomial |
Same as Poisson, handles overdispersion |
Same transforms apply |
# Standard Scaling
from sklearn.preprocessing import StandardScaler
X_scaled = StandardScaler().fit_transform(X)
# Log transform for positive skewed count data
X['log_visits'] = np.log1p(X['visits'])
# One-hot encoding for categorical
X = pd.get_dummies(X, columns=['device_type'])
# Offset for Poisson
model = sm.GLM(y, X, family=sm.families.Poisson(), offset=np.log(X['exposure']))
✅ Checklist Before Modeling
- [ ] Numeric features reviewed for skew or count-like structure
- [ ] Categorical variables encoded correctly (one-hot or ordinal)
- [ ] For Poisson/NB: offset column included (log of exposure)
- [ ] Scaled if using regularized logistic (L1/L2)
- [ ] Transforms documented and justified in EDA or notebook
💡 Tip
“When modeling probabilities or counts, log is your best friend — but only when used intentionally.”