U-CECE: A Universal Multi-Resolution Framework for Conceptual Counterfactual Explanations

2026-04-09

AI Summary

Existing concept-level counterfactual explanation methods struggle to balance expressive power and computational efficiency: atomic concepts are efficient but ignore relational structure, while graph-based representations preserve fidelity at the cost of NP-hard graph edit distance computations. This work proposes U-CECE, the first general, model-agnostic, multi-resolution counterfactual framework that adaptively switches among three levels—atomic concepts, relation sets, and structural graphs—and integrates both transductive GNN and inductive GAE paradigms. By introducing a scalable unsupervised graph autoencoder pathway, U-CECE circumvents expensive graph edit operations while preserving semantic fidelity. Experiments on datasets such as CUB and Visual Genome demonstrate that U-CECE effectively balances efficiency and expressiveness, with its structured counterfactuals achieving semantic equivalence or even superiority over exact graph-edit solutions in both human evaluations and assessments by large vision-language models.

Abstract

As AI models grow more complex, explainability is essential for building trust, yet concept-based counterfactual methods still face a trade-off between expressivity and efficiency. Representing underlying concepts as atomic sets is fast but misses relational context, whereas full graph representations are more faithful but require solving the NP-hard Graph Edit Distance (GED) problem. We propose U-CECE, a unified, model-agnostic multi-resolution framework for conceptual counterfactual explanations that adapts to data regime and compute budget. U-CECE spans three levels of expressivity: atomic concepts for broad explanations, relational sets-of-sets for simple interactions, and structural graphs for full semantic structure. At the structural level, both a precision-oriented transductive mode based on supervised Graph Neural Networks (GNNs) and a scalable inductive mode based on unsupervised graph autoencoders (GAEs) are supported. Experiments on the structurally divergent CUB and Visual Genome datasets characterize the efficiency-expressivity trade-off across levels, while human surveys and LVLM-based evaluation show that the retrieved structural counterfactuals are semantically equivalent to, and often preferred over, exact GED-based ground-truth explanations.

Problem

• counterfactual explanations
• concept-based explanation
• expressivity-efficiency trade-off
• Graph Edit Distance
• explainable AI

Innovation

• conceptual counterfactual explanations
• multi-resolution framework
• graph edit distance
• graph neural networks
• model-agnostic explainability