Enhancing Graph Learning Interpretability through Modulating Cluster Information Flow

Jiayi Yang, Wei Ye, Xin Sun, Rui Fan, Jungong Han

Abstract:

Interpretable graph learning is essential for scientific applications that depend on
learning models to extract reliable insights from graph-structured data. Recent efforts
to explain GNN predictions focus on identifying vital substructures, such as subgraphs.
However, existing approaches tend to misclassify the neighboring irrelevant nodes as
part of the vital subgraphs. To address this, we propose Cluster Information
Flow Graph Neural Networks (CIFlow-GNN), a built-in model-level method that
provides accurate interpretable subgraph explanations by modulating the cluster
information flow.  CIFlow-GNN incorporates two modules, i.e., the graph clustering
module and the cluster prototype module. The graph clustering module partitions the
nodes according to their connectivity in the graph topology and their similarity in cluster
features. Specifically, we introduce a cluster feature loss to regulate information flow at
the cluster level. We prove that the proposed cluster feature loss is a lower bound of
the InfoNCE loss. Optimizing the cluster feature loss reduces the mutual information
among clusters and achieves the modulation of cluster information flow. Subsequently,
the graph prototype module uses prototypes as a bridge to select important clusters as
vital subgraphs by integrating information across all graphs. To ensure accurate
correspondence between clusters and prototypes, we further modulate the cluster
information flow at the prototype level. Experimental studies on both synthetic and real world datasets demonstrate that our proposed CIFlow-GNN can identify vital
subgraphs effectively and efficiently.