Explainability
Attribution Methods
Techniques that assign importance scores to inputs/features
3 subcategories • 20 techniques
Gradient Based
Uses gradients/derivatives to compute feature importance (e.g., Integrated Gradients)
Model Specific
Leverages specific model architecture for attribution (e.g., Mean Decrease Impurity for trees)
Perturbation Based
Modifies inputs to measure impact on outputs (e.g., SHAP, Permutation Importance)
Techniques
20 techniques
| Goals | Models | Data Types | Description | |||
|---|---|---|---|---|---|---|
| SHapley Additive exPlanations | Algorithmic | Architecture/model Agnostic Requirements/black Box | Any | SHAP (SHapley Additive exPlanations) explains one prediction by sharing out the gap between that prediction and a... | ||
| Permutation Importance | Algorithmic | Architecture/model Agnostic Requirements/black Box | Tabular | Permutation Importance measures how much a trained model relies on each input feature. For each feature, the values in a... | ||
| Sobol Indices | Algorithmic | Architecture/model Agnostic Requirements/black Box | Tabular | Sobol indices show how much of the variation in a model's output can be traced to each of its inputs. The analyst sets a... | ||
| Local Interpretable Model-Agnostic Explanations | Algorithmic | Architecture/model Agnostic Requirements/black Box | Any | LIME (Local Interpretable Model-agnostic Explanations) explains individual predictions by approximating the complex... | ||
| Partial Dependence Plots | Visualisation | Architecture/model Agnostic Requirements/black Box | Tabular | Partial Dependence Plots show how changing one or two features affects a model's predictions on average. The technique... | ||
| Occlusion Sensitivity | Algorithmic | Architecture/model Agnostic Requirements/black Box | Image | Occlusion sensitivity shows which parts of an input a model's prediction is sensitive to. A region of the input, such as... | ||
| Contrastive Explanation Method | Algorithmic | Architecture/neural Networks Paradigm/discriminative +4 | Any | The Contrastive Explanation Method (CEM) explains one prediction of a classifier, such as a neural network, by... | ||
| Permutation Tests | Algorithmic | Architecture/model Agnostic Requirements/black Box | Any | A permutation test asks how likely an observed result would be if there were no real relationship in the data. The... | ||
| GNNExplainer | Algorithmic | Architecture/neural Networks Requirements/gradient Access +2 | Graph | GNNExplainer explains one prediction of a graph neural network, a model that passes information along the edges of a... | ||
| TimeSHAP | Algorithmic | Architecture/neural Networks/recurrent Requirements/black Box | Time-series | TimeSHAP explains one prediction of a model that reads a sequence of events, such as a recurrent neural network that... | ||
| Mean Decrease Impurity | Algorithmic | Architecture/tree Based Paradigm/supervised +1 | Tabular | Mean Decrease Impurity (MDI) scores how much each feature contributes to a tree-based model, such as a random forest, by... | ||
| Coefficient Magnitudes (in Linear Models) | Metric | Architecture/linear Models Paradigm/parametric +2 | Tabular | Coefficient Magnitudes explains a linear model by reading its fitted coefficients. Each coefficient shows how much the... | ||
| Layer-wise Relevance Propagation | Algorithmic | Requirements/model Internals Architecture/neural Networks +3 | Any | Layer-wise Relevance Propagation (LRP) explains a neural network's prediction by working backwards from the output to... | ||
| Contextual Decomposition | Algorithmic | Architecture/neural Networks/recurrent Requirements/white Box +1 | Text | Contextual Decomposition explains why an LSTM, a neural network that reads text one word at a time, made a particular... | ||
| Integrated Gradients | Algorithmic | Architecture/neural Networks Architecture/neural Networks/transformer +4 | Image Tabular Text | Integrated Gradients assigns input features scores for a prediction relative to a chosen reference input, called the... | ||
| DeepLIFT | Algorithmic | Requirements/model Internals Architecture/neural Networks +2 | Any | DeepLIFT (Deep Learning Important FeaTures) explains a single prediction of a neural network by dividing the difference... | ||
| Taylor Decomposition | Algorithmic | Requirements/model Internals Architecture/neural Networks +3 | Any | Taylor Decomposition explains one prediction of a neural network by splitting its output score into a contribution from... | ||
| Gradient Saliency | Algorithmic | Architecture/neural Networks Requirements/differentiable +2 | Image | Gradient saliency (also called vanilla gradient saliency) is the simplest gradient-based explanation for image... | ||
| Gradient-weighted Class Activation Mapping | Algorithmic | Architecture/neural Networks/convolutional Requirements/architecture Specific +3 | Image | Grad-CAM (Gradient-weighted Class Activation Mapping) draws a heatmap over an image showing which regions most raise a... | ||
| Influence Functions | Algorithmic | Architecture/linear Models Architecture/neural Networks +5 | Any | Influence functions estimate which training examples most affected a model's prediction on a particular test example.... |
Rows per page
Page 1 of 1