B-cos GNNs: Faithful Explanations through Dynamic Linearity

TL;DR

B-cos GNNs are an explainable graph neural network model that provides exact, efficient, and task-specific explanations through a single linear decomposition, replacing traditional non-linear functions.

Contribution

The paper introduces B-cos GNNs, a novel explainable GNN architecture that offers exact explanations with minimal computational overhead, unlike existing post-hoc methods.

Findings

Achieves state-of-the-art explainability with minimal accuracy loss.

Provides explanations in a single forward and backward pass.

Outperforms post-hoc baselines in explanation speed.

Abstract

We introduce B-cos GNNs, an inherently explainable class of graph neural networks whose predictions decompose exactly into per-node, per-feature contributions via a single input-dependent linear map. B-cos GNNs use linear (sum-based) aggregation and replace non-linear message and update functions with B-cos transforms. This induces meaningful, task-specific weight-input alignment that is directly accessible through the model's dynamic linearity. Instance-level explanations follow from a single forward and backward pass, requiring no auxiliary explainer, modified learning objective, or perturbation procedure. Instantiated as a GIN, our approach trades small losses in predictive accuracy for state-of-the-art explainability across diverse synthetic and real-world benchmarks, producing explanations orders of magnitude faster than post-hoc baselines.