InterpretationMethod

Enums · AiDotNet.Interpretability

Enumeration of interpretation methods supported by interpretable models.

For Beginners

These are the different explanation techniques you can use to understand why your model makes certain predictions. Each method has different strengths:

  • Model-Agnostic Methods (work with any model):

  • SHAP, LIME, PartialDependence, Counterfactual, DiCE, Anchor, FeatureImportance, FeatureInteraction, Occlusion, FeatureAblation

  • Neural Network Methods (require gradient access):

  • IntegratedGradients, DeepLIFT, DeepSHAP, GradientSHAP, GradCAM, LayerGradCAM, GuidedBackprop, GuidedGradCAM, NoiseTunnel

  • Tree-Based Methods (for tree models):

  • TreeSHAP

  • Concept-Based Methods (for high-level understanding):

  • TCAV

  • Training Data Attribution:

  • InfluenceFunctions

Enable methods through InterpretabilityOptions when configuring your model.

Fields

Field Summary
InterpretationMethod Anchor Anchor explanations for rule-based interpretations.
InterpretationMethod Counterfactual Counterfactual explanations to understand decision boundaries.
InterpretationMethod DeepLIFT DeepLIFT (Deep Learning Important FeaTures) for neural network attribution.
InterpretationMethod DeepSHAP DeepSHAP combining GradientSHAP with DeepLIFT for efficient neural network explanations.
InterpretationMethod DiCE DiCE (Diverse Counterfactual Explanations) for generating diverse what-if scenarios.
InterpretationMethod FeatureAblation Feature Ablation for attribution by removing/replacing features.
InterpretationMethod FeatureImportance Feature importance analysis using permutation importance.
InterpretationMethod FeatureInteraction Feature interaction analysis using Friedman’s H-statistic.
InterpretationMethod GradCAM Grad-CAM (Gradient-weighted Class Activation Mapping) for CNN visual explanations.
InterpretationMethod GradientSHAP GradientSHAP for gradient-based SHAP approximation on neural networks.
InterpretationMethod GuidedBackprop Guided Backpropagation for cleaner gradient visualizations.
InterpretationMethod GuidedGradCAM GuidedGradCAM combining GuidedBackprop with GradCAM for high-resolution class-specific visualizations.
InterpretationMethod InfluenceFunctions Influence Functions for training data attribution.
InterpretationMethod InputXGradient Input × Gradient attribution - multiplies input values by their gradients.
InterpretationMethod IntegratedGradients Integrated Gradients for neural network attribution.
InterpretationMethod LIME LIME (Local Interpretable Model-agnostic Explanations) for local explanations.
InterpretationMethod LayerAttribution Layer-level attribution for computing attributions at intermediate layers.
InterpretationMethod LayerGradCAM LayerGradCAM for class activation mapping at a specific network layer.
InterpretationMethod NeuronAttribution Neuron-level attribution for understanding individual neuron contributions.
InterpretationMethod NoiseTunnel NoiseTunnel (SmoothGrad) for noise-averaged gradient smoothing.
InterpretationMethod Occlusion Occlusion-based attribution by systematically hiding parts of input.
InterpretationMethod PartialDependence Partial dependence plots to show feature effects.
InterpretationMethod SHAP SHAP (SHapley Additive exPlanations) values for feature importance.
InterpretationMethod TCAV TCAV (Testing with Concept Activation Vectors) for concept-level explanations.
InterpretationMethod TracIn TracIn (Tracing Influence) for efficient training data attribution.
InterpretationMethod TreeSHAP TreeSHAP for exact SHAP values on tree-based models.
int value__