VERY EARLY STAGE Variable Importance Plot library
A very early stage, highly incomplete, proof-of-concept Python variable importance plotting library using single dispatch to mimic R's VIP library.
It's easy!
import numpy as np
import matplotlib.pyplot as plt
from sklearn import ensemble
from sklearn import datasets
from sklearn.utils import shuffle
from sklearn.metrics import mean_squared_error
import bigwig
boston = datasets.load_boston()
X, y = shuffle(boston.data, boston.target, random_state=13)
X = X.astype(np.float32)
offset = int(X.shape[0] * 0.9)
X_train, y_train = X[:offset], y[:offset]
X_test, y_test = X[offset:], y[offset:]
params = {'n_estimators': 500, 'max_depth': 4, 'min_samples_split': 2,
'learning_rate': 0.01, 'loss': 'ls'}
clf = ensemble.GradientBoostingRegressor(**params)
clf.fit(X_train, y_train)
# defaults:
bigwig.vip(clf, boston.feature_names, relative=True, num_features=10, bar=True,
horizontal=True, color="lightgrey", fill="lightgrey")
# Vertical: bigwig.vip(clf, boston.feature_names, horizontal=False)
Currently only with GB and RF regressions from Scikit Learn.
This project has been set up using PyScaffold 3.1. For details and usage information on PyScaffold see https://pyscaffold.org/.