A Multiscale Graph Convolutional Network Using Hierarchical Clustering
This incremental work addresses the need for better modeling of hierarchical structures in networks, with potential applications in domains like molecular property prediction and protein interface prediction.
The authors tackled the problem of underutilizing hierarchical topology in networks by proposing a multiscale graph convolutional network that uses hierarchical clustering to learn representations from fine to coarse scales, achieving competitive performance on a benchmark citation network.
The information contained in hierarchical topology, intrinsic to many networks, is currently underutilised. A novel architecture is explored which exploits this information through a multiscale decomposition. A dendrogram is produced by a Girvan-Newman hierarchical clustering algorithm. It is segmented and fed through graph convolutional layers, allowing the architecture to learn multiple scale latent space representations of the network, from fine to coarse grained. The architecture is tested on a benchmark citation network, demonstrating competitive performance. Given the abundance of hierarchical networks, possible applications include quantum molecular property prediction, protein interface prediction and multiscale computational substrates for partial differential equations.