Attributed Modularity Loss for an Improved Graph Clustering

Abstract

Graph Neural Network (GNN) based architectures have been widely used for performing clustering tasks on attributed graphs. In this context, a commonly used loss function employed to train the GNNs is modularity (𝑄𝑚𝑜𝑑 ) − a topology-based measure that is also used for evaluating the cluster quality. However, the 𝑄𝑚𝑜𝑑 loss used in previous works consider only the graph structure, and ignores the nodal attributes. This drawback may have its implications on the overall quality of clusters obtained. In order to address this issue, we propose a novel modularity measure, called Attributed-Modularity (𝑄𝑎𝑡𝑡𝑟 ), that integrates both the structural information and nodal features. Experimentally, we validated our approach across five real world datasets, showing highly improved cluster accuracies.