shap_feature_importance.py

shap_feature_importance.py

import shap
import numpy as np
import pandas as pd

categoric_features = tuple(['FEATURE1', 'FEATURE2', 'ETC'])

def avaliar_importancias_features(modelo_treinado, X):
    explainer = shap.Explainer(modelo_treinado)
    shap_values = explainer.shap_values(X)

    vals = np.abs(shap_values).mean(0)
    feature_importance = pd.DataFrame(list(zip(X.columns, sum(vals))), columns=['Feature', 'Peso'])
    feature_importance.sort_values(by=['Peso'], ascending=False,inplace=True)
    feat_imp_dict = feature_importance.to_dict(orient='records')
    
    feat_imp_dict_final = {}

    # Somar importâncias de features categóricas
    # Utilizado pois o oneHotEncoding feito pelo pd.get_dummies
    # Adiciona um underline ao nome de cada feature categórica, seguido do nome da categoria
    for fi in feat_imp_dict:
        if fi['Feature'].startswith(categoric_features):
            real_feature_name = fi['Feature'][:fi['Feature'].rfind('_')]

            if real_feature_name in feat_imp_dict_final:
                feat_imp_dict_final[real_feature_name] = feat_imp_dict_final[real_feature_name] + fi['Peso']
            else:
                feat_imp_dict_final[real_feature_name] = fi['Peso']
        else:
            feat_imp_dict_final[fi['Feature']] = fi['Peso']

    # Retorna os mais importantes primeiro
    return dict(sorted(feat_imp_dict_final.items(), key=lambda item: item[1], reverse=True))