Paper accepted at GDDL @ NeurIPS 2026

20 September 2026

We are excited to share our new paper, “FlowCF: Sparse Counterfactual Explanations for Mixed-Type Tabular Data using Flow Matching,” accepted at the Geometric Distributional Deep Learning Workshop at NeurIPS 2026.

This work is authored by PhD researcher Emmanouil Panagiotou and Prof. Eirini Ntoutsi.

Counterfactual explanations make AI decisions easier to understand by showing which input features would need to change to produce a different outcome. However, existing methods often modify too many features or propose unnecessarily large changes, limiting how understandable and actionable their explanations are.

In this paper, we introduce FlowCF, a model-agnostic generative method that frames counterfactual generation as sparse transport between classes using flow matching.

Key takeaways:

  • Sparse explanations. FlowCF uses a learned gating mechanism to identify and modify only the features necessary to change a prediction.

  • Support for mixed-type data. A new mixed flow operator handles numerical and categorical features within a unified framework.

  • Efficient and model-agnostic generation. Once trained, FlowCF generates counterfactuals without gradients or queries to the underlying black-box model.

  • Strong empirical results. Across six benchmark datasets, FlowCF changed only 29% of numerical features, compared with 89% for the strongest baseline, while reducing the size of the changes by 70%.

These results show that flow matching can produce counterfactual explanations that are more concise, actionable, and faithful to the original instance.