Introduction to
Machine Learning Explainability

Part II

Kacper Sokol

Topics

• Classification of Explanations
• Explanation Modalities
• Examples of Explanations
• Data Explainability
• Transparent Modelling
• Post-hoc Explainability

Classification of Explanations

O1 Explanation Family

• associations between antecedent and consequent
• contrasts and differences
• causal mechanisms

Associations Between Antecedent and Consequent

• feature importance
• feature attribution / influence
• rules

• exemplars (prototypes & criticisms)

Contrasts and Differences

• (non-causal) counterfactuals
  i.e., contrastive statements
• prototypes & criticisms

Causal Mechanisms

• causal counterfactuals
• causal chains
• full causal model

Explanation Modalities

O2 Explanatory Medium

• (statistical / numerical) summarisation
• visualisation
• textualisation
• formal argumentation

O4 Explanation Domain

Original domain

Transformed domain

(O3 System Interaction & U4 Interactiveness]

Provided within a static or interactive protocol

• interactive interface
• interactive explanation

Examples of Explanations

Permutation Feature Importance

https://www.kaggle.com/code/dansbecker/permutation-importance

Individual Conditional Expectation &
Partial Dependence

FACE Counterfactuals

Poyiadzi, Sokol, Santos-Rodriguez, De Bie and Flach, 2020. FACE: Feasible and actionable counterfactual explanations

RuleFit

https://christophm.github.io/interpretable-ml-book/rulefit.html

Data Explainability

• Data as an (implicit) model
• Data summarisation and description
• Exemplars, prototypes and criticisms
• Dimensionality reduction (e.g., t-SNE)

Transparent Modelling

• Rule lists and sets
• Linear models
• Decision trees
• \(k\)-nearest neighbours and \(k\)-means

Post-hoc Explainability

Understandable model of the relation between inputs and outputs

• SHAP
• LIME