Simulate RF Categorical Variable Encoding Bias

rf_bias.py

import random
import sys

import numpy as np
import matplotlib.pyplot as plt
from sklearn.ensemble import RandomForestClassifier

N_SAMPLES = 1000
N_TREES = 100
MAX_CATEGORIES = 32
N_SIMS = 100

def generate_onehot(n_samples, max_categories):
    n_features = 0
    variable_indices = [0]
    for i in xrange(2, max_categories + 1):
        n_features += i
        variable_indices.append(n_features)
    
    features = np.zeros((n_samples, n_features))
    labels = np.zeros((n_samples))
    for r in xrange(n_samples):
        labels[r] = random.randint(0, 1)

        for c in xrange(n_features):
            features[r, c] = random.randint(0, 1)

    return features, labels, variable_indices

def generate_stacked(n_samples, max_categories):
    n_features = max_categories - 1

    features = np.zeros((n_samples, n_features))
    labels = np.zeros((n_samples))
    for r in xrange(n_samples):
        labels[r] = random.randint(0, 1)

        for c in xrange(n_features):
            n_categories = c + 2
            features[r, c] = random.randint(0, n_categories - 1)

    return features, labels

def plot_variable_importances(flname, variable_importances, title):
    plt.clf()
    plt.boxplot(x=variable_importances)
    plt.xlabel("Variable", fontsize=16)
    plt.ylabel("Gini Importance", fontsize=16)
    plt.title(title, fontsize=18)
    plt.savefig(flname, DPI=200)

if __name__ == "__main__":
    # burn in
    for i in xrange(100):
        random.random()

    variable_importances = [[] for i in xrange(MAX_CATEGORIES - 1)]
    for i in xrange(N_SIMS):
        print "Round", i
        X, y, variable_indices = generate_onehot(N_SAMPLES, MAX_CATEGORIES)
        rf = RandomForestClassifier(n_estimators = N_TREES)
        rf.fit(X, y)
        feature_importances = rf.feature_importances_

        for i, (start, end) in enumerate(zip(variable_indices[:-1], variable_indices[1:])):
            variable_importances[i].extend(feature_importances[start:end])

    plot_variable_importances(sys.argv[1], variable_importances, "One-Hot Encoded")

    variable_importances = [[] for i in xrange(MAX_CATEGORIES - 1)]
    for i in xrange(N_SIMS):
        print "Round", i
        X, y = generate_stacked(N_SAMPLES, MAX_CATEGORIES)
        rf = RandomForestClassifier(n_estimators = N_TREES)
        rf.fit(X, y)
        feature_importances = rf.feature_importances_

        for i in xrange(MAX_CATEGORIES - 1):
            variable_importances[i].append(feature_importances[i])

    plot_variable_importances(sys.argv[2], variable_importances, "Integer Encoded")

If you have converted 31 features to 527 features using one-hot encoding then after feature_importances = rf.feature_importances_ , you got importance for only 31 original features which shown in the plot and not for 527 features those are actually used for fit the model. Thank you for sharing your knowledge! I look forward to hearing from you.

Indeed, sklearn requires OneHotEncoder of categorical variables for RandomForest otherwise they categorical variables will be considered as continuous. My interpretation is that here @rnowling demonstrates the consequence of treating the categorical variables as continuous. This code and the article about "Feature Importance Bias" is related to specific implementation of RF in sklearn. I would not expect the same result in other implementations such as H2O, R or Spark, because theoretically decision trees perfectly handle categorical variables.