GradientSHAPExplainer<T>

Models & Types · AiDotNet.Interpretability.Explainers

GradientSHAP explainer - a faster approximation of SHAP using gradients.

Implements: IGPUAcceleratedExplainer<T>, IGlobalAttributionExplainer<T>, ILocalExplainer<T, GradientSHAPExplanation<T>>

For Beginners

GradientSHAP combines ideas from Integrated Gradients and SHAP to create a faster method for explaining neural networks.

How it works:

  1. Sample random baselines from your background data
  2. For each baseline, compute something like Integrated Gradients
  3. Average the results to get SHAP-like values

Comparison with other methods:

  • KernelSHAP: Model-agnostic but slow (doesn’t use gradients)
  • DeepSHAP: Fast but requires access to layer activations
  • GradientSHAP: Good balance - uses gradients but doesn’t need internal model access

Why use GradientSHAP?

  • Much faster than KernelSHAP for neural networks
  • Only requires gradient computation (not layer access)
  • Approximates SHAP values reasonably well
  • Better than plain gradients (uses baselines for context)

When to use:

  • You have a neural network
  • KernelSHAP is too slow
  • You don’t have access to model internals (for DeepSHAP)

Constructors

Constructor Summary
GradientSHAPExplainer(Func<Vector<T>, Vector<T>> predictFunction, Func<Vector<T>, int, Vector<T>> gradientFunction, Matrix<T> backgroundData, int numSamples, int numSteps, bool addNoise, double noiseStdDev, string[] featureNames, int? randomState) Initializes a new GradientSHAP explainer.

Properties

Property Summary
bool IsGPUAccelerated { get; }
bool SupportsGlobalExplanations { get; }
bool SupportsLocalExplanations { get; }
string MethodName { get; }

Methods

Method Summary
GradientSHAPExplanation<T> Explain(Vector<T> instance) Computes GradientSHAP attributions for an input.
GradientSHAPExplanation<T> Explain(Vector<T> instance, int outputIndex) Computes GradientSHAP attributions for a specific output.
GradientSHAPExplanation<T>[] ExplainBatch(Matrix<T> instances)
Vector<T> ComputeGlobalAttributions(Matrix<T> data)
void SetGPUHelper(GPUExplainerHelper<T> helper)