PACE neuro-symbolic framework generates realistic, feasible counterfactual explanations

Counterfactual explanations—showing minimal changes to flip a model’s prediction—often recommend unrealistic actions like "reduce your age" or "get a PhD overnight." The new PACE framework tackles this by separating prediction from reasoning: a neural model handles classification, while a symbolic layer (using Answer Set Programming) encodes real-world constraints. This modular design ensures that generated counterfactuals respect domain knowledge, such as preserving immutable features (e.g., age, race) and only suggesting feasible modifications (e.g., increasing work hours, changing occupation).

In a case study on the Adult Income dataset, PACE combined a multilayer perceptron with ASP rules that limited changes to education, occupation, and working hours. Results showed a clear trade-off between counterfactual validity (achieving the opposite prediction) and plausibility (realistic recommendations). PACE’s symbolic constraints produced explanations that, while sometimes needing larger changes, were far more actionable and interpretable than those from purely neural baselines.

The framework is model-agnostic, meaning it can work with any classifier, and its explicit constraint encoding makes it adaptable to domains like healthcare, finance, and HR, where unrealistic explanations can cause real harm. By bridging neural and symbolic AI, PACE moves toward transparent decision support that users can trust and act upon.