ksfeatureselector 0.2.0

A robust and flexible Python package designed for selecting the most discriminatory features in both **binary and multi-class classification problems** using the Kolmogorov-Smirnov (K-S) test. It provides advanced options for handling multi-class scenarios and aggregating p-values.

KSFeatureSelector is a robust and flexible Python package designed for selecting the most discriminatory features in both binary and multi-class classification problems using the Kolmogorov-Smirnov (K-S) test. It provides advanced options for handling multi-class scenarios and aggregating p-values.

Features

• Uses the K-S test to rank features by their ability to separate classes.

• Handles target variables with more than two categories (up to 10 classes internally).

• Flexible Comparison Strategies:

  • pairwise: Performs K-S tests between every unique pair of classes.

  • one-vs-rest: Compares each class against all other classes combined.

• Multiple P-Value Aggregation Methods:

  • fisher: Uses Fisher’s combined probability test (default, generally recommended).

  • min: Takes the minimum p-value from all comparisons for a feature.

  • max: Takes the maximum p-value from all comparisons for a feature.

• Scikit-learn Style API: Offers a class-based interface (KSFeatureSelector with fit, transform) for seamless integration into machine learning pipelines.

• Convenience Function: Provides a simple select_ks_features wrapper for quick, one-off feature selection.

• Robust Validation & Warnings: Includes comprehensive input validation and issues UserWarning for data quality issues, such as categories with too few observations or insufficient samples for K-S tests.

• Pure Python: Built using pandas, scipy, and numpy.

Installation

pip install ksfeatureselector

For local installation:

pip install -e .

Usage

from ksfeatureselector import select_ks_features

significant_features = select_ks_features(
    df, x_cols, y_var,
    top_p=0.01,
    aggregation_method='one-vs-rest',
    p_value_aggregation_method='min'
)
print(f"Significant features (one-vs-rest, min p-value <= 0.01): {significant_features}")

# Example 3: Select top 3 features using 'pairwise' comparison
# and 'max' p-value aggregation
top_3_features_max_agg = select_ks_features(
    df, x_cols, y_var,
    top_n=3,
    aggregation_method='pairwise',
    p_value_aggregation_method='max'
)
print(f"Top 3 features (pairwise, max p-value): {top_3_features_max_agg}")

Arguments

• df (pd.DataFrame): The input DataFrame containing feature columns and the binary target column.

• x_cols (List[str]): A list of column names in df representing the features you want to evaluate.

• y_var (str): The name of the column in df representing the binary target variable (0/1 or similar).

• top_p (float, optional): If provided, only features with a K-S test p-value less than top_p will be selected.

• top_n (int, optional): If provided, the top n features with the lowest p-values will be selected.

  Note

  You can use either top_p or top_n, or both. If both are given, the function will apply top_p first, and then take the top n from that filtered list.

License

MIT License

Author

V Subrahmanya Raghu Ram Kishore Parupudi Email: [e-mail]