Mutual Information for Genuine Hypothesis Testing (MIGHT)

An example using FeatureImportanceForestClassifier for nonparametric multivariate hypothesis test, on simulated datasets. Here, we present a simulation of how MIGHT is used to test the hypothesis that a “feature set is important for predicting the target”. This is a generalization of the framework presented in [1].

We simulate a dataset with 1000 features, 500 samples, and a binary class target variable. Within each feature set, there is 500 features associated with one feature set, and another 500 features associated with another feature set. One could think of these for example as different datasets collected on the same patient in a biomedical setting. The first feature set (X) is strongly correlated with the target, and the second feature set (W) is weakly correlated with the target (y). Here, we are testing the null hypothesis:

• H0: I(X; y) - I(X, W; y) = 0

• HA: I(X; y) - I(X, W; y) < 0 indicating that there is more mutual information with

  respect to y

where I is mutual information. For example, this could be true in the following settings, where X is our informative feature set and W is our uninformative feature set.

• W    X -> y: here W is completely disconnected from X and y.

• W -> X -> y: here W is d-separated from y given X.

• W <- X -> y: here W is d-separated from y given X.

We then use MIGHT to test the hypothesis that the first feature set is important for predicting the target, and the second feature set is not important for predicting the target. We use FeatureImportanceForestClassifier.

import numpy as np
from scipy.special import expit

from sktree import HonestForestClassifier
from sktree.stats import FeatureImportanceForestClassifier
from sktree.tree import DecisionTreeClassifier

seed = 12345
rng = np.random.default_rng(seed)

Simulate data

We simulate the two feature sets, and the target variable. We then combine them into a single dataset to perform hypothesis testing.

n_samples = 2000
n_features_set = 20
mean = 1.0
sigma = 2.0
beta = 5.0

unimportant_mean = 0.0
unimportant_sigma = 4.5

# first sample the informative features, and then the uniformative features
X_important = rng.normal(loc=mean, scale=sigma, size=(n_samples, 10))
X_important = np.hstack(
    [
        X_important,
        rng.normal(
            loc=unimportant_mean, scale=unimportant_sigma, size=(n_samples, n_features_set - 10)
        ),
    ]
)

X_unimportant = rng.normal(
    loc=unimportant_mean, scale=unimportant_sigma, size=(n_samples, n_features_set)
)
X = np.hstack([X_important, X_unimportant])

# simulate the binary target variable
y = rng.binomial(n=1, p=expit(beta * X_important[:, :10].sum(axis=1)), size=n_samples)

Perform hypothesis testing using Mutual Information

Here, we use FeatureImportanceForestClassifier to perform the hypothesis test. The test statistic is computed by comparing the metric (i.e. mutual information) estimated between two forests. One forest is trained on the original dataset, and one forest is trained on a permuted dataset, where the rows of the covariate_index columns are shuffled randomly.

The null distribution is then estimated in an efficient manner using the framework of [1]. The sample evaluations of each forest (i.e. the posteriors) are sampled randomly n_repeats times to generate a null distribution. The pvalue is then computed as the proportion of samples in the null distribution that are less than the observed test statistic.

n_estimators = 100
max_features = "sqrt"
test_size = 0.2
n_repeats = 1000
n_jobs = -1

est = FeatureImportanceForestClassifier(
    estimator=HonestForestClassifier(
        n_estimators=n_estimators,
        max_features=max_features,
        tree_estimator=DecisionTreeClassifier(),
        random_state=seed,
        honest_fraction=0.25,
        n_jobs=n_jobs,
    ),
    random_state=seed,
    test_size=test_size,
    permute_per_tree=False,
    sample_dataset_per_tree=False,
)

print(
    f"Permutation per tree: {est.permute_per_tree} and sampling dataset per tree: "
    f"{est.sample_dataset_per_tree}"
)
# we test for the first feature set, which is important and thus should return a pvalue < 0.05
stat, pvalue = est.test(
    X, y, covariate_index=np.arange(n_features_set, dtype=int), metric="mi", n_repeats=n_repeats
)
print(f"Estimated MI difference: {stat} with Pvalue: {pvalue}")

# we test for the second feature set, which is unimportant and thus should return a pvalue > 0.05
stat, pvalue = est.test(
    X,
    y,
    covariate_index=np.arange(n_features_set, dtype=int) + n_features_set,
    metric="mi",
    n_repeats=n_repeats,
)
print(f"Estimated MI difference: {stat} with Pvalue: {pvalue}")

Permutation per tree: False and sampling dataset per tree: False
Estimated MI difference: -0.014835853389523923 with Pvalue: 0.002997002997002997
Estimated MI difference: 0.002754835563470198 with Pvalue: 0.7822177822177823

References

[1] (1,2)

Tim Coleman, Wei Peng, and Lucas Mentch. Scalable and efficient hypothesis testing with random forests. The Journal of Machine Learning Research, 23(1):7679–7713, 2022.