PACE Framework Generates Plausible AI Counterfactual Explanations

Explaining machine learning predictions often involves generating counterfactuals, which are minimal input changes that would alter a model's decision. However, many existing methods frequently produce unrealistic or impractical recommendations because they lack mechanisms to integrate real-world domain knowledge and intervention constraints. PACE, a new neuro-symbolic framework, addresses this by separating prediction from reasoning. It uses a neural predictive model for classification and a symbolic reasoning layer to enforce domain-specific rules during the counterfactual generation process. This modular approach ensures that the generated explanations are not only interpretable but also plausible and actionable within a given context. The framework is model-agnostic and adaptable, making it suitable for various domains requiring realistic decision support. A case study on the Adult Income dataset demonstrated that PACE, by combining a multilayer perceptron with Answer Set Programming rules, successfully generated explanations that better satisfied feasibility requirements, highlighting the value of neuro-symbolic AI in explainable AI.