Towards Point Cloud Counterfactual Explanations

Abstract

Counterfactual explanations, engineered alterations of classifier inputs that change predictions, are widely used for interpreting classifier decisions. In this work, we extend counterfactual generation to point cloud data, addressing the challenges posed by its unstructured nature. We introduce hierarchical modeling to enhance counterfactual learning, specifically the ability to identify and manipulate key semantic traits influencing classifier predictions. Through visual and quantitative evaluations, we demonstrate the effectiveness of our method in generating counterfactuals that successfully reverse classifier predictions.

Publication
2025 33rd European Signal Processing Conference (EUSIPCO)
Nicolas Vercheval
Doctoral researcher

My current research interests include quaternion neural networks, geometric deep learning, and mesh processing.