HumanCompatible.Explain Documentation

This documentation contains the explainability methods developed for the AutoFair project. These tools are designed to provide insight into machine learning models through counterfactual explanations, with a focus on fairness and transparency.

Included Methods

• GLANCE: A versatile and adaptive framework for generating global counterfactual explanations. These explanations are expressed as actions that offer recourse to large population subgroups. The framework aims to provide explanations and insights, ensuring that the actions benefit as many individuals as possible.

• FACTS: An efficient, model-agnostic, highly parameterizable, and explainable framework for evaluating subgroup fairness through counterfactual explanations. FACTS facilitates in-depth analysis of fairness across different subgroups by simulating alternative decision outcomes.

• LiCE: A method of finding high-quality Counterfactual Explanations. Utilizing a Sum-Product Network to estimate the plausibility (i.e., likelihood) of the counterfactual sample. We are able to find globally optimal (i.e., closest/most likely) counterfactuals, while satisfying various constraints on the data (actionability, data constraints, causal relations…).

• FCX: Feasible Counterfactual Explanations (FCX) is a novel framework that generates realistic and low-cost counterfactuals by enforcing both hard feasibility constraints provided by domain experts and soft causal constraints inferred from data.

Explore the documentation below for details on the implementation, usage, and customization of these methods.

Contents

• FACTS: Fairness Aware Counterfactual for Subgroups package
  • FACTS
  • Rule Mining & Filtering
  • Metrics
  • Optimization & Utilities
• GLANCE: Global Actions in a Nutshell for Counterfactual Explainability
  • Global Counterfactual Methods
  • Clustering Method Wrappers
  • Local Counterfactual Methods
  • Utility Functions
• LiCE: Likely Counterfactual Explanations
  • LiCE
  • Sum-Product Network wrapper
  • Data handling
• FCX: Finding Feasible Counterfactual Explanation
  • FCX Unary Constraint Training (Adult)
  • FCX Binary Constraint Training (Adult)
  • FCX BlackBox Model training (Adult)
  • FCX Unary Evaluation (Adult)
  • FCX Binary Evaluation (Adult)