Algorithmic Concept-based Explainable Reasoning

TL;DR

This paper introduces concept-bottleneck GNNs that enhance explainability by learning and extracting propositional formulas and global concepts, while maintaining competitive performance on graph tasks.

Contribution

It proposes a novel GNN modification for concept-based explanations, enabling interpretable reasoning without explicit supervision.

Findings

Accurately learns concepts and propositional formulas.

Achieves performance comparable to state-of-the-art models.

Derives global graph concepts without supervision.

Abstract

Recent research on graph neural network (GNN) models successfully applied GNNs to classical graph algorithms and combinatorial optimisation problems. This has numerous benefits, such as allowing applications of algorithms when preconditions are not satisfied, or reusing learned models when sufficient training data is not available or can't be generated. Unfortunately, a key hindrance of these approaches is their lack of explainability, since GNNs are black-box models that cannot be interpreted directly. In this work, we address this limitation by applying existing work on concept-based explanations to GNN models. We introduce concept-bottleneck GNNs, which rely on a modification to the GNN readout mechanism. Using three case studies we demonstrate that: (i) our proposed model is capable of accurately learning concepts and extracting propositional formulas based on the learned concepts for each target class; (ii) our concept-based GNN models achieve comparative performance with state-of-the-art models; (iii) we can derive global graph concepts, without explicitly providing any supervision on graph-level concepts.