mdi_calculation.py

import pandas as pd
import numpy as np

total_mdi_scores = pd.DataFrame(columns=features)

for i, (train, test) in enumerate(cv.split()):
  rf.fit(X[train], y[train], sample_weights=weights[train])
  
  # Create a DataFrame with the total loss per feature (columns) per tree (rows)
  mdi_scores = [i: tree.feature_importances_ for i, tree in enumerate(rf.estimators_)]
  mdi_scores = pd.DataFrame.from_dict(mdi_scores, orient="index")
  mdi_scores.columns = features
  mdi_scores = mdi_scores.replace(0, np.nan)
  total_mdi_scores = total_mdi_scores.append(mdi_scores)

# Calculate the mean and std of the loss reduction for each feature
mdi_scores = pd.concat({'mean': mdi_scores.mean(), 'std': mdi_scores.std()*mdi_scores.shape[0]**-.5}, axis=1)

# Divide the mean loss reduction of each feature by the total loss reduction accross all features
mdi_scores /= mdi_scores['mean'].sum()