Faithful Explanations for Graph Classification using Logic @ ECML-PKDD 2025
Today I had the pleasure of presenting our paper “Faithful Explanations for Graph Classification using Logic” at ECML PKDD 2025, in collaboration with Marc Plantevit and Céline Robardet.
In this work, we introduce LogiX, a method that combines Graph Neural Networks with a transparent logic layer to produce explanations that are faithful to the model’s reasoning and easy to interpret. Unlike many post-hoc techniques, our approach directly models the decision process, which lets us identify the most relevant nodes and extract global logic rules.
Across synthetic and molecular datasets, LogiX delivers explanations that are more faithful, sparser, and more stable compared to existing methods, helping bridge the gap between black-box predictions and human-understandable reasoning.
