skboost 0.2.0

Relational features, adaptive search, and confidence evaluation tools for scikit-learn

skboost

skboost is a lightweight Python library designed to boost your models — whether trees, linear models, neural networks, or anything else. It provides relational feature engineering, adaptive hyperparameter search, and confidence-aware evaluation tools that enhance model performance and interpretability.

Installation

pip install skboost

Features

Preprocessing

RelationalFeaturesTransformer - Boost model performance with compact relational features

from skboost.preprocessing import RelationalFeaturesTransformer

transformer = RelationalFeaturesTransformer(direction='larger')
X_transformed = transformer.fit_transform(X)
# For each feature, finds indices and distances to next larger/smaller values
# Adds O(N) features instead of O(N²) pairwise combinations
# Helps models learn inter-feature relationships
# Useful for: computer vision, ranking tasks, any data with meaningful ordering

DualScalerTransformer - Boost classification with class-specific scaling

from skboost.preprocessing import DualScalerTransformer

transformer = DualScalerTransformer()
X_scaled = transformer.fit_transform(X)
# Scales features separately for different target classes
# Useful for: imbalanced datasets, multi-class problems with different distributions

Hyperparameter Tuning

zoom_search_cv - Boost efficiency with adaptive hyperparameter search

from skboost.tuning import zoom_search_cv

best_params, best_score = zoom_search_cv(
    estimator, X, y,
    param_grid={'n_estimators': [50, 100, 150], 'max_depth': [3, 6, 9]},
    n_iter=3, cv=5
)
# Starts with 3 values per parameter, iteratively zooms around best region
# Works for both numeric and categorical hyperparameters
# Useful for: faster optimization than exhaustive grid search

Model Evaluation

confidence_report - Boost reliability with confidence-aware evaluation

from skboost.evaluation import confidence_report, plot_confidence_report

reports = confidence_report(y_true, y_proba, thresholds=[0.5, 0.7, 0.9])
plot_confidence_report(reports)
# Shows precision/recall/f1 at different confidence thresholds per class
# Visualize which predictions your model is reliable on
# Useful for: production deployment decisions, finding usable subsets, model monitoring

Additional Tools

GroupDiffTransformer - Sequential feature engineering within groups

from skboost.preprocessing import GroupDiffTransformer

transformer = GroupDiffTransformer(key_col='user_id')
X_transformed = transformer.fit_transform(X)
# Adds: difference from previous row, difference from first row per group

GroupValueCountsTransformer - Value frequency features within groups

from skboost.preprocessing import GroupValueCountsTransformer

transformer = GroupValueCountsTransformer(group_col='session_id', value_col='action')
X_transformed = transformer.fit_transform(X)
# Adds: raw counts and normalized counts per group

Quick Example

import numpy as np
from sklearn.ensemble import RandomForestClassifier
from sklearn.datasets import make_classification
from skboost.preprocessing import RelationalFeaturesTransformer
from skboost.tuning import zoom_search_cv
from skboost.evaluation import confidence_report, plot_confidence_report

# Generate data
X, y = make_classification(n_samples=500, n_classes=3, n_informative=5, random_state=42)

# Add relational features
transformer = RelationalFeaturesTransformer(direction='larger')
X_boosted = transformer.fit_transform(X)

# Adaptive hyperparameter search
param_grid = {'n_estimators': [50, 100, 150], 'max_depth': [3, 6, 9]}
clf = RandomForestClassifier(random_state=42)
best_params, best_score = zoom_search_cv(clf, X_boosted, y, param_grid, n_iter=3)

# Train with best params and evaluate confidence
clf.set_params(**best_params)
clf.fit(X_boosted, y)
y_proba = clf.predict_proba(X_boosted)

# Confidence-stratified evaluation
reports = confidence_report(y, y_proba, thresholds=[0.5, 0.7, 0.9])
plot_confidence_report(reports)

Testing

pytest tests/

See tests/ directory for usage examples in test form.

License

MIT