Parallelizing Node-Level Explainability in Graph Neural Networks

TL;DR

This work tackles the inefficiency of node-level explainability in Graph Neural Networks by introducing a parallelization framework based on graph partitioning, augmented with cluster-border neighbor reconstruction to preserve explanation correctness. It leverages METIS to partition the graph into clusters, reconstructs border neighborhoods either fully or via dropout to balance memory and fidelity, and processes one node per cluster in parallel to compute explanations. The approach provides theoretical complexity bounds and demonstrates substantial runtime speedups on standard benchmarks while enabling production-scale, transparent explanations for large GNNs. The technique is agnostic to the underlying explainability method and offers a tunable memory–accuracy trade-off, with practical implications for real-time interpretability in industrial settings.

Abstract

Graph Neural Networks (GNNs) have demonstrated remarkable performance in a wide range of tasks, such as node classification, link prediction, and graph classification, by exploiting the structural information in graph-structured data. However, in node classification, computing node-level explainability becomes extremely time-consuming as the size of the graph increases, while batching strategies often degrade explanation quality. This paper introduces a novel approach to parallelizing node-level explainability in GNNs through graph partitioning. By decomposing the graph into disjoint subgraphs, we enable parallel computation of explainability for node neighbors, significantly improving the scalability and efficiency without affecting the correctness of the results, provided sufficient memory is available. For scenarios where memory is limited, we further propose a dropout-based reconstruction mechanism that offers a controllable trade-off between memory usage and explanation fidelity. Experimental results on real-world datasets demonstrate substantial speedups, enabling scalable and transparent explainability for large-scale GNN models.

Figures (8)

• Figure 1: Example of node importance scores computed using two explainability techniques: (a) Saliency Map simonyanDeepConvolutionalNetworks2014a, (b) GNNExplainer gnnexplainer. Scores are shown for node classification on the Cora dataset using a Graph Convolutional Network (GCN) gcn. The red node is the target being classified, and the surrounding nodes are the neighbors considered. Importance values are in the $\left[0,1\right]$ range, with 1 indicating maximum importance.
• Figure 2: Forward and backward passes of a 1-hop GNN. (a) Black lines represent the forward flow of information; (b) red lines indicate the backward flow during gradient computation.
• Figure 3: Overview of the proposed methodology. Red nodes and edges represent the reconstructed neighbors, while blue nodes denote, for each batch, the selected nodes from each cluster for which explainability is executed in parallel.
• Figure 4: Example of graph partitioning into clusters, first step of the proposed parallelization process. (a) Original graph; (b) cluster division of the graph, where each color represents a different cluster. The edges 3-5 and 7-6 have been broken due to the cluster creation process.
• Figure 5: Example of cluster-border neighbor reconstruction applied to the partition of \ref{['fig: division']} using a 1-hop GNN. Reconstructed nodes are shown in gray. (a) Full reconstruction method; (b) dropout reconstruction, where the reconstructed edges are drawn with dashed lines to indicate there is a probability of restoring each edge.