18 Sep 2025 · Bilbao, Spain
The central question behind this work is how to produce explanations for graph classification that stay faithful to the actual decision process. That is an important distinction because a model can sound convincing while still hiding the mechanisms that really drive the prediction. In this paper, we use a logic-oriented layer to keep the explanation connected to the computation itself.
The result is a model that can describe its decision in a more compact and inspectable way. Instead of treating interpretability as an afterthought, the system integrates a transparent structure that makes the relevant nodes and relations easier to track. That is especially useful when the goal is to understand why a graph belongs to a class, not just to obtain a score.
What improves in practice
In our experiments, the logic-based approach produced explanations that were easier to compare across datasets and more stable than many alternatives. That stability matters because graph explanations can otherwise vary substantially from one run to the next, which makes them difficult to use as part of a scientific discussion.
Why faithful matters more than pretty
A visually appealing explanation is not necessarily a reliable one. The main point of LogiX is to make the explanation line up with the model behavior, even when that means the result is more constrained. For me, that trade-off is worth making when the aim is to support scientific reasoning and not just to produce a nice visualization.
View the LogiX code and see the paper entry in Publications.