fselect 1.0.4

Feature Selection for Clustering

Feature Selection for Clustering: fselect

A fast and scalable implementation of A-RANK algorithm as proposed by Dash, M. and Liu, H. in their paper "Feature Selection for Clustering" for selecting features from a dataset using an entropy measure using fast python libraries: numpy, pandas and scikit-learn.

Getting Started

Install the package:

pip install fselect

Import the main function:

from fselect import rank_features  

Prepare a dataframe with normalized continuous features:

import pandas as pd

df = pd.DataFrame({
    'feature1': [...],
    'feature2': [...],    
    [...]
})

Rank the features:

ranked_df = rank_features(df)  

The returned dataframe `ranked_df` contains columns: "rank", "feature", "entropy" sorted by entropy.

Usage

The main parameters:

• dataframe: pd.DataFrame - Input dataframe with continuous normalized features
• remove_correlated_columns: bool (optional) - Whether to remove highly correlated columns before ranking
• correlation_threshold: float (optional) - Correlation threshold to determine correlated columns (default 0.999)

Remove correlated columns first

ranked_df = rank_features(df, remove_correlated_columns=True)  

Custom correlation threshold

ranked_df = rank_features(df, remove_correlated_columns=True, correlation_threshold=0.95) 

Algorithm

The entropy calculation is based on the equations defined in the ARANK paper. It calculates a similarity matrix of the dataframe and computes entropy from the same.