FlowCF: Sparse Counterfactual Explanations for Mixed-Type Tabular Data using Flow Matching

FlowCF achieves significantly sparser explanations: The method changes only 29% of numerical features compared to 89% for the best existing baseline, meaning it identifies fewer input changes needed to explain why a model made a particular decision—making explanations more practical and easier for users to understand.

Explanations require smaller, more realistic changes: When modifications are necessary, FlowCF requires 70% smaller adjustments to feature values than competing methods, producing counterfactual explanations that are more feasible and believable in real-world contexts.

Works with mixed data types without requiring model access: Unlike some explanation methods that need direct access to a model's internal workings, FlowCF is "model-agnostic," meaning it can explain any type of AI system while handling both numerical and categorical data—a common scenario in real business applications like credit decisions or medical diagnoses.

Addresses a practical gap in explainability: The paper identifies that existing explanation methods struggle particularly with numerical features (like income or age), and FlowCF specifically improves performance in this area, making it more useful for explaining decisions in domains where numerical data dominates.

Balances multiple explanation qualities simultaneously: The method maintains competitive performance across multiple desirable properties of explanations (sparsity, proximity, and other factors), rather than excelling at one while sacrificing others—important for building trustworthy AI systems in practice.