Shapley Values

The primary model-agnostic implementation of Shapley values for machine learning. SHAP unifies six existing interpretability methods by framing feature importance as an additive feature attribution problem. It computes Shapley values efficiently using techniques like KernelSHAP (model-agnostic) and TreeSHAP (optimized for tree-based models). The output is a linear formula where each feature's contribution sums to the difference between the model's prediction and the average prediction.

A local surrogate method that explains individual predictions by perturbing the input and fitting a sparse linear model to the resulting predictions. Unlike Shapley values, LIME does not guarantee a fair distribution of the prediction among features. It relies on heuristic sampling around the instance, making it faster but less theoretically grounded. LIME defines locality via a kernel, while Shapley values define it via the coalitional game theory framework.

An axiomatic attribution method designed specifically for differentiable models like deep neural networks. It satisfies the completeness axiom (attributions sum to the prediction difference) by integrating gradients along a straight-line path from a baseline input to the actual input. This path integral is the continuous analog of the discrete Shapley value summation over coalitions. It avoids the exponential computational cost of exact Shapley values for high-dimensional inputs.

A framework that applies Shapley values to training data valuation rather than feature attribution. It quantifies the contribution of each individual training data point to the final model performance. Computing exact Data Shapley values requires retraining the model on all 2^N subsets of data, which is intractable. Practical approximations use Monte Carlo sampling with truncated permutation sampling to estimate the marginal contribution of each datum.

A method that extends Shapley values from local (single prediction) to global feature importance. SAGE applies the Shapley value formula to the model's overall loss function, measuring how much each feature reduces predictive error when included. It accounts for complex interactions that permutation importance misses. SAGE uses a sampling-based approximation that leverages the model's ability to handle missing features via marginalization.

An adaptation that replaces the observational conditional distribution used in standard SHAP with a causal interventional distribution. Standard Shapley values can assign importance to features that are merely correlated with the outcome, not causally related. Causal Shapley values use a Structural Causal Model (SCM) to simulate interventions, breaking spurious correlations. This ensures feature importance reflects true causal drivers, critical for high-stakes decisions.