Statistical Context Engineering

Statistical Context Engineering (SCE) is a feature engineering methodology that enriches tabular datasets with hierarchical statistical context, improving regression model performance.

Key Features

• Leakage-safe: Uses cross-fitting (out-of-fold aggregation) to prevent target leakage
• Hierarchical: Supports multi-level categorical hierarchies with automatic backoff
• scikit-learn compatible: Drop-in transformer for ML pipelines
• Auto-detection: Automatically identifies categorical columns

Installation

pip install stat-context

Quick Example

from sce import StatisticalContextEngine, ContextConfig

# Configure the engine
config = ContextConfig(
    target_col="price",
    use_cross_fitting=True  # Prevents leakage
)

# Create and fit the engine
engine = StatisticalContextEngine(config)
enriched_df = engine.fit_transform(train_df)

# Now enriched_df has additional statistical context features

How It Works

SCE computes group-level statistics (mean, std, count, etc.) for categorical columns, creating features that capture the "context" of each observation. The cross-fitting approach ensures these features don't leak target information.

For example, for a property in a specific neighborhood: - neighborhood_price_mean: Average price in that neighborhood - neighborhood_price_std: Price variation in that neighborhood
- neighborhood_price_count: Number of samples in that neighborhood

Documentation

• Getting Started
• API Reference
• Experiments

Citation

If you use SCE in your research, please cite:

@software{stachowicz2025sce,
  author = {Stachowicz, Mateusz and Halkiewicz, Stanisław},
  title = {Statistical Context Engineering: Hierarchical Feature Enrichment for Regression Models},
  year = {2025},
  url = {https://github.com/joint-hubs/sce}
}