SHAP

SHAP explains a prediction as a sum of individual feature contributions. The model's output is decomposed into a base value (the average prediction) plus the additive Shapley value for each feature.

• Linear Decomposition: The final prediction equals the sum of all feature attributions plus a constant base value.
• Local Accuracy: The explanation perfectly matches the original model output for the specific instance being explained.
• Contrastive: Each value represents the change in prediction when conditioning on that feature, relative to the average.

SHAP is rooted in cooperative game theory, where each feature is a 'player' in a coalition. The Shapley value is the unique solution that satisfies four axioms of fair profit distribution.

• Efficiency: The sum of all contributions equals the total prediction minus the average.
• Symmetry: Features that contribute identically receive identical values.
• Dummy: A feature that never changes the prediction receives a value of zero.
• Additivity: For ensemble models, the SHAP value of the ensemble is the average of individual model SHAP values.

KernelSHAP is a model-agnostic method that estimates Shapley values using a specially weighted linear regression. It treats the model as a black box, requiring only the ability to evaluate the model on different input subsets.

• Weighting Kernel: Uses the Shapley kernel to weight coalitions, giving higher importance to small and large coalitions.
• Sampling Strategy: Approximates the exponential number of coalitions by sampling, trading computational cost for estimation accuracy.
• Background Dataset: Requires a representative sample of data to serve as the baseline for missing feature values during coalition evaluation.

TreeSHAP computes exact Shapley values for tree-based models in polynomial time, avoiding the exponential complexity of naive calculation. It leverages the internal structure of decision trees for precise, fast attribution.

• Algorithm Complexity: Runs in O(TLD²) time where T is trees, L is leaves, and D is depth, making it feasible for large ensembles like XGBoost and LightGBM.
• Feature Interactions: Directly captures pairwise interaction effects, decomposing predictions into main effects and interaction terms.
• No Sampling Error: Unlike KernelSHAP, TreeSHAP yields deterministic, exact values without approximation noise.

DeepSHAP combines DeepLIFT's backpropagation rules with Shapley value principles to efficiently approximate SHAP values for deep learning models. It linearizes network components to propagate attributions rapidly.

• Linear Composition: Treats common operations like ReLU, max-pooling, and convolutions as linearizable components.
• Reference Values: Uses a background sample to define the 'off' state for each neuron, analogous to the baseline in Integrated Gradients.
• Speed Advantage: Orders of magnitude faster than KernelSHAP for deep networks, enabling practical use on large medical imaging models.

SHAP is the only additive feature attribution method that satisfies the consistency property: if a model changes so that a feature's marginal contribution increases or stays the same, that feature's attributed importance should not decrease.

• Theoretical Uniqueness: Proven to be the unique solution within the additive attribution class that satisfies local accuracy, missingness, and consistency.
• Comparison to Other Methods: Unlike LIME, which uses a heuristic kernel, SHAP's weighting is theoretically justified. Unlike simple gradients, SHAP satisfies completeness.
• Regulatory Alignment: This theoretical grounding makes SHAP particularly suitable for regulatory submissions where explanation fidelity must be demonstrable.

A high-speed approximation algorithm that combines DeepLIFT's rules with Shapley value calculations. Designed specifically for deep learning models, it efficiently estimates feature attributions by leveraging the compositional architecture of neural networks.

• Built as a SHAP extension for deep models
• Uses linear composition rules to propagate attributions
• Dramatically faster than kernel-based SHAP for large networks
• Default explainer for deep models in the SHAP library

Local Interpretable Model-agnostic Explanations takes a fundamentally different approach from SHAP. It approximates the model locally with an interpretable surrogate model rather than computing exact Shapley values.

• Model-agnostic: works on any black-box classifier
• Generates perturbed samples around a prediction
• Fits a simple linear model to the local neighborhood
• Does not guarantee the completeness axiom that SHAP satisfies

An axiomatic attribution method that satisfies sensitivity and implementation invariance. It computes importance by integrating gradients along a straight-line path from a baseline to the input.

• Satisfies the completeness axiom (attributions sum to prediction difference)
• Requires a meaningful baseline (often black image or zeros)
• Directly applicable to any differentiable model
• Often used alongside SHAP for validation in regulatory submissions

Produces coarse localization maps by using gradient signals flowing into the final convolutional layer. Unlike SHAP's per-pixel attribution, Grad-CAM highlights spatial regions most influential for a class decision.

• Specifically designed for CNN architectures
• Generates class-discriminative visualizations
• Commonly used in radiology AI for lesion localization
• Complements SHAP by providing spatial context to feature importance

A quantitative metric that evaluates explanation quality by measuring how well attributed importance correlates with actual model behavior when features are perturbed.

• Perturb top-k attributed features and measure output change
• Higher correlation indicates more faithful explanation
• Used to benchmark SHAP against other methods
• Critical for regulatory explainability validation in SaMD