ForwardStepwiseFeatureSelection aims to select the best features or the subset of features in machine learning tasks according to corresponding score with other incredible packages like numpy, pandas and sklearn.
# Install ForwardStepwiseFeatureSelection
!pip install ForwardStepwiseFeatureSelection
# Import dependenices
from ForwardStepwiseFeatureSelection import ForwardStepwiseFeatureSelection
from sklearn.model_selection import train_test_split
from sklearn.preprocessing import StandardScaler
from sklearn.ensemble import RandomForestRegressor
import pandas as pd
# Read dataframe
insurance = pd.read_csv('insurance.csv')
# Label Encoding
for col in ['sex', 'smoker', 'region']:
insurance[col].replace(insurance[col].unique(), range(insurance[col].nunique()), inplace=True)
X = insurance.drop('charges', axis=1)
y = insurance['charges']
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.33, random_state=42)
# Scale our data
scaler = StandardScaler()
X_train, X_test = (pd.DataFrame(scaler.fit_transform(df), columns=df.columns) for df in [X_train, X_test])
X_train, X_test, y_train, y_test = train_test_split(X, y)
# Instantiate the estimator
rfc = RandomForestRegressor()
# Instantiate ForwardStepwiseFeatureSelection
fsfs = ForwardStepwiseFeatureSelection(estimators=rfc, cv=3, scoring='neg_mean_absolute_error', mode=None, verbose=1, tolerance=3)
# Start feature selection
fsfs.fit(X_train, y_train)
print(fsfs.best_subsets)
>> {'RandomForestRegressor': ['smoker', 'age', 'bmi', 'children', 'region']}
This package is inspired by: PyData DC 2016 | A Practical Guide to Dimensionality Reduction Vishal Patel October 8, 2016
- Examples: https://github.com/HindyDS/ForwardStepwiseFeatureSelection/tree/main/examples
- Email: [email protected]
- Source code: https://github.com/HindyDS/ForwardStepwiseFeatureSelection/tree/main/ForwardStepwiseFeatureSelection
- Bug reports: https://github.com/HindyDS/ForwardStepwiseFeatureSelection/issues
It requires at least six arguments to run:
- estimators: machine learning model
- X (array): features space
- y (array): target
- cv (int): number of folds in a (Stratified)KFold
- scoring (str): see https://scikit-learn.org/stable/modules/model_evaluation.html
Optional arguments:
- mode (string): None or 'ts'. If 'ts' (Time Series) than it will change to walk forward cross validation.
- max_trial (int): number of trials that you wanted FSFS to stop searching
- tolerance (int): how many times FSFS can fail to find better subset of features
- least_gain (int): threshold of scoring metrics gain in fraction
- max_feats (int): maximum number of features
- prior (list): starting point for FSFS to search, must be corresponds to the columns of X
- exclusions (nested list): if the new selected feature is in one of the particular subpool (list in the nested list), then the features in that particular subpool with no longer be avalible to form any new subset in the following trials
- n_jobs (int): Number of jobs to run in parallel.
- n_digit (int): Decimal places for scoring
- verbose (int): Level of verbosity of FSFS
If you have any ideas for this packge please don't hesitate to bring forward!