PACE: A Neuro-Symbolic Framework for Plausible and Actionable Counterfactual Explanations

Counterfactual explanations can be unrealistic: Existing methods for explaining AI decisions often suggest changes that are impractical or impossible in real life—for example, recommending someone instantly change their education level—because they lack built-in safeguards for what's actually feasible.

Domain knowledge matters for trustworthy AI: PACE combines machine learning with symbolic reasoning (human-understandable rules) to ensure explanations respect real-world constraints and domain expertise, making AI recommendations more credible and actionable for decision-makers.

The framework distinguishes between changeable and unchangeable factors: By explicitly modeling which attributes can be modified (like occupation or working hours) versus which cannot (like age or past education), the system produces fairer and more realistic explanations, particularly important in high-stakes domains like lending or hiring.

There's a trade-off between accuracy and practicality: The paper identifies a fundamental tension between generating explanations that reliably change a model's prediction and keeping those explanations realistic and feasible—a key consideration when deploying AI systems in real-world applications.

Model-agnostic design increases applicability: PACE works with different types of AI models and can be adapted across various industries, suggesting this approach to responsible AI explanation could be broadly useful beyond the income prediction case study presented.