Using scikit to determine contributions of each feature to a specific class prediction


Being more knowledgable about ML today than I was 2.5 years ago, I will now say this approach only works for highly linear decision problems. If you carelessly apply it to a non-linear problem you will have trouble.

Example: Imagine a feature for which neither very large nor very small values predict a class, but values in some intermediate interval do. That could be water intake to predict dehydration. But water intake probably interacts with salt intake, as eating more salt allows for a greater water intake. Now you have an interaction between two non-linear features. The decision boundary meanders around your feature-space to model this non-linearity and to ask only how much one of the features influences the risk of dehydration is simply ignorant. It is not the right question.

Alternative: Another, more meaningful, question you could ask is: If I didn't have this information (if I left out this feature) how much would my prediction of a given label suffer? To do this you simply leave out a feature, train a model and look at how much precision and recall drops for each of your classes. It still informs about feature importance, but it makes no assumptions about linearity.

Below is the old answer.

I worked through a similar problem a while back and posted the same question on Cross Validated. The short answer is that there is no implementation in sklearn that does all of what you want.

However, what you are trying to achieve is really quite simple, and can be done by multiplying the average standardised mean value of each feature split on each class, with the corresponding model._feature_importances array element. You can write a simple function that standardises your dataset, computes the mean of each feature split across class predictions, and does element-wise multiplication with the model._feature_importances array. The greater the absolute resulting values are, the more important the features will be to their predicted class, and better yet, the sign will tell you if it is small or large values that are important.

Here's a super simple implementation that takes a datamatrix X, a list of predictions Y and an array of feature importances, and outputs a JSON describing importance of each feature to each class.

def class_feature_importance(X, Y, feature_importances):
    N, M = X.shape
    X = scale(X)

    out = {}
    for c in set(Y):
        out[c] = dict(
            zip(range(N), np.mean(X[Y==c, :], axis=0)*feature_importances)

    return out


import numpy as np
import json
from sklearn.preprocessing import scale

X = np.array([[ 2,  2,  2,  0,  3, -1],
              [ 2,  1,  2, -1,  2,  1],
              [ 0, -3,  0,  1, -2,  0],
              [-1, -1,  1,  1, -1, -1],
              [-1,  0,  0,  2, -3,  1],
              [ 2,  2,  2,  0,  3,  0]], dtype=float)

Y = np.array([0, 0, 1, 1, 1, 0])
feature_importances = np.array([0.1, 0.2, 0.3, 0.2, 0.1, 0.1])
#feature_importances = model._feature_importances

result = class_feature_importance(X, Y, feature_importances)

print json.dumps(result,indent=4)

    "0": {
        "0": 0.097014250014533204, 
        "1": 0.16932975630904751, 
        "2": 0.27854300726557774, 
        "3": -0.17407765595569782, 
        "4": 0.0961523947640823, 
        "5": 0.0
    "1": {
        "0": -0.097014250014533177, 
        "1": -0.16932975630904754, 
        "2": -0.27854300726557779, 
        "3": 0.17407765595569782, 
        "4": -0.0961523947640823, 
        "5": 0.0

The first level of keys in result are class labels, and the second level of keys are column-indices, i.e. feature-indices. Recall that large absolute values corresponds to importance, and the sign tells you whether it's small (possibly negative) or large values that matter.

This is modified from the docs

from sklearn import datasets
from sklearn.ensemble import ExtraTreesClassifier

iris = datasets.load_iris()  #sample data
X, y =,

model = ExtraTreesClassifier(n_estimators=10000, n_jobs=-1, random_state=0)
model.fit_transform(X,y) # fit the dataset to your model

I think feature_importances_ is what you're looking for:

In [13]: model.feature_importances_
Out[13]: array([ 0.09523045,  0.05767901,  0.40150422,  0.44558631])


Maybe I misunderstood the first time (pre-bounty), sorry, this may be more along the lines of what you are looking for. There is a python library called treeinterpreter that produces the information I think you are looking for. You'll have to use the basic DecisionTreeClassifer (or Regressor). Following along from this blog post, you can discretely access the feature contributions in the prediction of each instance:

from sklearn import datasets
from sklearn.cross_validation import train_test_split
from sklearn.tree import DecisionTreeClassifier

from treeinterpreter import treeinterpreter as ti

iris = datasets.load_iris()  #sample data
X, y =,
#split into training and test 
X_train, X_test, y_train, y_test = train_test_split( 
    X, y, test_size=0.33, random_state=0)

# fit the model on the training set
model = DecisionTreeClassifier(random_state=0),y_train)

I'll just iterate through each sample in X_test for illustrative purposes, this almost exactly mimics the blog post above:

for test_sample in range(len(X_test)):
    prediction, bias, contributions = ti.predict(model, X_test[test_sample].reshape(1,4))
    print "Class Prediction", prediction
    print "Bias (trainset prior)", bias

    # now extract contributions for each instance
    for c, feature in zip(contributions[0], iris.feature_names):
        print feature, c

    print '\n'

The first iteration of the loop yields:

Class Prediction [[ 0.  0.  1.]]
Bias (trainset prior) [[ 0.34  0.31  0.35]]
sepal length (cm) [ 0.  0.  0.]
sepal width (cm) [ 0.  0.  0.]
petal length (cm) [ 0.         -0.43939394  0.43939394]
petal width (cm) [-0.34        0.12939394  0.21060606]

Interpreting this output, it seems as though petal length and petal width were the most important contributors to the prediction of third class (for the first sample). Hope this helps.

The paper "Why Should I Trust You?": Explaining the Predictions of Any Classifier was submitted 9 days after this question, providing an algorithm for a general solution to this problem! :-)

In short, it is called LIME for "local interpretable model-agnostic explanations", and works by fitting a simpler, local model around the prediction(s) you want to understand.

What's more, they have made a python implementation ( with pretty detailed examples on how to use it with sklearn. For instance this one is on two-class random forest problem on text data, and this one is on continuous and categorical features. They are all to be found via the README on github.

The authors had a very productive year in 2016 concerning this field, so if you like reading papers, here's a starter:

  • Programs as Black-Box Explanations
  • Nothing Else Matters: Model-Agnostic Explanations By Identifying Prediction Invariance
  • Model-Agnostic Interpretability of Machine Learning