FeatureAblationExplainer<T>

Models & Types · AiDotNet.Interpretability.Explainers

Feature Ablation explainer for understanding feature importance by removal.

Implements: IGPUAcceleratedExplainer<T>, IGlobalAttributionExplainer<T>, IGlobalExplainer<T, GlobalFeatureAblationResult<T>>, ILocalExplainer<T, FeatureAblationExplanation<T>>

For Beginners

Feature Ablation is a simple way to understand feature importance. The idea is: replace each feature with a “baseline” value and see how the prediction changes.

How it works:

  1. For each feature (or group of features):
  2. Replace the feature with a baseline (e.g., zero, mean, or reference value)
  3. Measure how much the prediction changes
  4. Features that cause large changes are important

Difference from Occlusion:

  • Occlusion: Slides a window over spatial data (images)
  • Feature Ablation: Works on individual features (tabular data, any modality)

Key advantage: You can group features and ablate them together:

  • All color channels for an image
  • All words in a sentence segment
  • All related features in tabular data

Use cases:

  • Understanding which features drive predictions in tabular data
  • Grouping related features (e.g., all “income” related columns)
  • Debugging models by finding unexpected important features

Constructors

Constructor Summary
FeatureAblationExplainer(Func<Tensor<T>, Tensor<T>> tensorPredictFunction, int[][] featureGroups, string[] featureNames, bool perturbEachSeparately) Initializes a Feature Ablation explainer for tensor inputs.
FeatureAblationExplainer(Func<Vector<T>, Vector<T>> predictFunction, Vector<T> baseline, int[][] featureGroups, string[] featureNames, bool perturbEachSeparately) Initializes a new Feature Ablation explainer.

Properties

Property Summary
bool IsGPUAccelerated { get; }
bool SupportsGlobalExplanations { get; } Gets whether this explainer supports global explanations.
bool SupportsLocalExplanations { get; } Gets whether this explainer supports local explanations.
string MethodName { get; } Gets the method name.

Methods

Method Summary
FeatureAblationExplanation<T> Explain(Vector<T> instance) Explains a single input (ILocalExplainer interface).
FeatureAblationExplanation<T> ExplainLocal(Vector<T> input, int? targetClass) Explains a single input by ablating features.
FeatureAblationExplanation<T>[] ExplainBatch(Matrix<T> instances)
GlobalFeatureAblationResult<T> ExplainGlobal(Matrix<T> data) Computes global feature importance across a dataset.
GlobalFeatureAblationResult<T> ExplainGlobal(Matrix<T> data, int? targetClass) Computes global feature importance across a dataset for a specific class.
Vector<T> ComputeGlobalAttributions(Matrix<T> data) Explains a batch of inputs (ILocalExplainer interface).
void SetGPUHelper(GPUExplainerHelper<T> helper)