jcorpac Feature Engineering Studio
Feature engineering is often where the most significant model performance gains are found.
Core Techniques
- Encoding: One-Hot (low cardinality) vs. Target Encoding (high cardinality).
- Scaling: StandardScaler (Gaussian data) vs. RobustScaler (data with outliers).
- Binning: Converting numerical values into categorical ranges (e.g., Age -> Age Group).
- Interaction Terms: Multiplying features together to capture combined effects.
Safety: Avoiding Data Leakage
- CRITICAL: Never calculate scaling parameters (mean, std) on the test set.
- Fit your encoders/scalers ONLY on the training data and use them to transform both train and test.
Scikit-Learn Pipeline Integration
Implement custom transformers using BaseEstimator and TransformerMixin for seamless integration into sklearn.pipeline.Pipeline.