An Improvement of Cluster-GCN with constraints

Abstract

Cluster-GCN is one of the effective methods in studying the scalability of Graph Neural Networks. The idea of this approach is to use METIS community detection algorithm to split the graph into several sub-graphs that are small enough to fit into a common GPU. However, METIS algorithm still has some limitations. Therefore, this project proposes Leiden algorithm as an alternative, which was scientifically published 21 years after METIS’s and claimed to be powerful in identifying communities in networks. However, the common feature of community detection algorithms makes nodes in the same community tend to be similar. For that reason, this project also proposes to add constraints such as minimum/maximum community size and overlapping communities to increase community diversity, thereby improving performance of Cluster-GCN by 0.98% ROC-AUC score on a single 8GB GPU device.