Explainable AI: SHAP, LIME, and Feature Attribution

Module: Machine Learning | Difficulty: Advanced

Shapley Values

SHAP (KernelSHAP)

LIME

Anchors

import numpy as np
from itertools import combinations

def shapley_values(model, X, background, n_samples=100):
    n_features = X.shape[1]
    phi = np.zeros(n_features)
    for i in range(n_features):
        for _ in range(n_samples):
            perm = np.random.permutation(n_features)
            pos = np.where(perm == i)[0][0]
            before = perm[:pos]
            after = perm[pos+1:]
            # Marginal contribution
            f_before = model(np.mean(background[:, before], axis=0, keepdims=True))
            f_before_plus = model(np.mean(background[:, before + [i]], axis=0, keepdims=True))
            phi[i] += (f_before_plus - f_before) / n_samples
    return phi

Research Insight: SHAP values are the only feature attribution method that satisfies efficiency, symmetry, dummy, and linearity axioms. The computational complexity is exponential in the number of features, making exact computation intractable for large models.