feature-selection-lofo

LOFO

Leave One Feature Out (LOFO) is on of the most powerful techniques for feature selection.

This repository contains the implementation of LOFO in Python and can be used with any model of the followings:

• Any Scikit-Learn model.
• Any TensorFlow/Keras model.
• LightGBM.
• CatBoost.
• XGBoost.

Usage

• Install the package:

pip install feature-selection-lofo

• Import lofo

from feature_selection_lofo import lofo

lofo.LOFO(X, Y, 
          model, 
          cv, 
          metric, 
          direction, 
          fit_params=None, 
          predict_type='predict', 
          return_bad_feats=False, 
          groups=None,
          is_keras_model=False)

Args
X	Pandas DataFrame, input features to the model (predictors).
Y	array_like, target/label feature.
model	object, the model class (e.g. sklearn.linear_model.LinearRegression()).
cv	object, sklearn cross validatoin object (e.g. sklearn.model_selection.KFold(n_splits=5, shuffle=True, random_state=0)).
metric	object, metric to use during search (e.g. sklearn.metrics.roc_auc_score).
direction	string, direction of optimization ('max' or 'min').
fit_params	string, parameters to use for fitting (e.g. "{'X': x_train, 'y': y_train}") . Defaults to "{'X': x_train, 'y': y_train}".
predict_type	string, ('predict' or 'predict_proba'). Defaults to 'predict'.
return_bad_feats	boolean, whether to return a list of bad features. Defaults to False.
groups	array_like, used with StratifiedGroupKFold. Defaults to None.
is_keras_model	boolean, whether the model passed is Keras model. Defaults to False.

Returns
A Pandas DataFrame with harmful features removed.
If return_bad_feats is set to True, it returns a list of the harmful features.

• Import the needed libraries for your model, cross-validation, etc

Scikit-Learn Model Example

import warnings
import numpy as np
from sklearn.metrics import roc_auc_score
from sklearn.model_selection import StratifiedKFold
from sklearn.linear_model import LogisticRegression

• Define the paramters

# shutdown warning messages
warnings.filterwarnings('ignore')

X = train_df.iloc[:, :-1]
Y = train_df.iloc[:, -1]
model = LogisticRegression()
cv = StratifiedKFold(n_splits=5, shuffle=True, random_state=0)
metric = roc_auc_score
direction = 'max'
fit_params = "{'X': x_train, 'y': y_train}"
predict_type = 'predict_proba'
return_bad_feats = True
groups = None
is_keras_model = False

• Define the LOFO object and call it

lofo_object = lofo.LOFO(X, Y, model, cv, metric, direction, fit_params, 
                        predict_type, return_bad_feats, groups, is_keras_model)

clean_X, bad_feats = lofo_object()

clean_X: is the dataset containing the useful features only.

bad_feats: are the harmful or useless features.

LightGBM Model Example

import warnings
import numpy as np
from sklearn.metrics import roc_auc_score
from sklearn.model_selection import StratifiedKFold
import lightgbm as lgbm

• Define the paramters

# shutdown warning messages
warnings.filterwarnings('ignore')

X = train_df.iloc[:, :-1]
Y = train_df.iloc[:, -1]
model= lgbm.LGBMClassifier(
                       objective='binary',
                       metric='auc',
                       subsample=0.7,
                       learning_rate=0.03,
                       n_estimators=100,
                       n_jobs=-1)
cv = StratifiedKFold(n_splits=5, shuffle=True, random_state=0)
metric = roc_auc_score
direction = 'max'
fit_params = "{'X': x_train, 'y': y_train, 'eval_set': [(x_valid,y_valid)], 'verbose': 0}"
predict_type = 'predict_proba'
return_bad_feats = True
groups = None
is_keras_model = False

• Define the LOFO object and call it

lofo_object = lofo.LOFO(X, Y, model, cv, metric, direction, fit_params, 
                        predict_type, return_bad_feats, groups, is_keras_model)
clean_X, bad_feats = lofo_object()

TensorFlow/Keras Model Example

import warnings
import numpy as np
from sklearn.metrics import roc_auc_score
from sklearn.model_selection import StratifiedKFold
import tensorflow as tf
from tensorflow.keras import layers

• Construct the model

def nn_model():
    inputs = layers.Input(shape=X.shape[-1],)
    x = layers.Dense(256, activation='relu')(inputs)
    x = layers.Dense(64, activation='relu')(x)
    output = layers.Dense(1, activation='sigmoid')(x)
    
    model = tf.keras.Model(inputs=inputs, outputs=output)
    model.compile(loss='binary_crossentropy',
              optimizer='adam',)
    
    return model

• Define the paramters

# shutdown warning messages
warnings.filterwarnings('ignore')

X = train_df.iloc[:, :-1]
Y = train_df.iloc[:, -1]

tf.keras.backend.clear_session()
model = nn_model()

cv = StratifiedKFold(n_splits=5, shuffle=True, random_state=0)
metric = roc_auc_score
direction = 'max'
fit_params = "{'x': x_train, 'y': y_train, 'validation_data': (x_valid, y_valid), 'epochs': 10, 'batch_size': 256, 'verbose': 0}"
predict_type = 'predict'
return_bad_feats = True
groups = None
is_keras_model = True

• Define the LOFO object and call it

lofo_object = lofo.LOFO(X, Y, model, cv, metric, direction, fit_params, 
                        predict_type, return_bad_feats, groups, is_keras_model)

clean_X, bad_feats = lofo_object()