1 Dec 2025 · San Diego, USA
The main idea behind LogiX-GIN is simple to state and harder to build well: can a graph neural network expose its reasoning in a form that looks like logic? That question is useful because graph learning often works well while staying difficult to explain. LogiX-GIN answers that gap by shaping the architecture so that the learned representation can be translated into readable rules.
What makes the approach interesting is that the logic view is not an external interpretation step. It is connected to the network design itself, so the explanation follows the model instead of being approximated after the fact. That gives us a more direct way to inspect which structures in the graph matter, how they contribute to the class decision, and what kind of symbolic description the trained system actually supports.
Why this direction is useful
Self-explainable models are attractive when the goal is not only to maximize accuracy, but to preserve a traceable decision process. For graph problems, that matters because the user often wants to understand which substructures, motifs, or relations influenced the prediction. LogiX-GIN tries to make that evidence easier to read without falling back to a separate post-hoc explanation layer.
From scores to rules
A useful side effect of the logic-oriented design is that the same model can be discussed in both machine learning and symbolic terms. That is a practical advantage when collaborating with researchers who want a compact rule summary, or when comparing explanations across different datasets and training runs.
View the code on GitHub and check the matching entry in Publications for the paper link.