Counterfactual Metarules for Local and Global Recourse

The document introduces T-CREx, a model-agnostic method for local and global counterfactual explanation, which leverages tree-based surrogate models to learn human-readable rules and metarules, providing both a global analysis of model behavior and diverse recourse options for users, with superior aggregate performance over existing rule-based baselines on a range of counterfactual explanation desiderata.

Abstract

We introduce T-CREx, a novel model-agnostic method for local and global counterfactual explanation (CE), which summarises recourse options for both individuals and groups in the form of human-readable rules. It leverages tree-based surrogate models to learn the counterfactual rules, alongside 'metarules' denoting their regions of optimality, providing both a global analysis of model behaviour and diverse recourse options for users. Experiments indicate that T-CREx achieves superior aggregate performance over existing rule-based baselines on a range of CE desiderata, while being orders of magnitude faster to run.

Method

The method presented in this document, T-CREx, is a model-agnostic approach for explaining the outputs of black box AI models and summarizing recourse options for individuals and groups using human-interpretable rules. It leverages tree-based surrogate models to learn counterfactual rules and metarules, enabling both local and global counterfactual explanation. The method aims to provide diverse sets of counterfactual explanations for single inputs, enhancing the trust and understanding of non-expert users, and demonstrates strong performance relative to baselines on a range of counterfactual explanation desiderata.

Main Finding

The main finding of this document is the introduction of a model-agnostic method, T-CREx, for explaining the outputs of black box AI models and summarizing recourse options for individuals and groups using human-interpretable rules, demonstrating strong performance in providing diverse counterfactual explanations for single inputs and enabling both local and global counterfactual explanation, with superior aggregate performance over existing rule-based baselines on a range of counterfactual explanation desiderata.

Conclusion

The conclusion of this document is that the T-CREx method, a model-agnostic approach for explaining the outputs of black box AI models and summarizing recourse options for individuals and groups using human-interpretable rules, demonstrates strong performance in providing diverse counterfactual explanations for single inputs and enabling both local and global counterfactual explanation, with superior aggregate performance over existing rule-based baselines on a range of counterfactual explanation desiderata. The method's efficacy has been demonstrated on benchmark classification and regression datasets, and while it currently has limitations related to specific cost functions and domain-specific actionability constraints, it shows potential for generalization to handle actionability and alternative costs without major changes. Additionally, the document highlights the importance of further theoretical and empirical investigation into hyperparameters and the trade-offs between accuracy, feasibility, sparsity, and complexity in the context of counterfactual explanations.

Keywords

Counterfactual, Explanation, Feasibility, Sparsity, Complexity, Consistency, Accuracy, Runtime, XGBoost, Neural Network, Hyperparameter, Regression, Classification, Desiderata, Metarules, Rules, Group-level, Recourse, Counterfactual fairness, AI models, Human-interpretable.