Node Classification with Multi-hop Graph Convolutional Network

Various graph neural network architectures have emerged, each focusing on different facets of graphs, such as ensuring scalability, preserving local structures, and retaining structural and feature information within the node vectors. However, incorporating global structures into the latent representation is frequently overlooked despite being a central focus of many random-walk-based embedding techniques predating the advent of graph neural networks. Ensuring the preservation of global structural information enriches the node representations to tackle the node classification task effectively. We introduce an innovative multi-hopped graph convolutional network designed for graph-structured data to address this gap. This architecture leverages adjacent node information and incorporates distanced neighborhood information to preserve global structure in the latent representations. We propose a method for aggregating multiple hopped features and devise a straightforward yet effective architecture. Empirical validation of our theoretical findings on various graph classification benchmarks showcases that our model attains state-of-the-art and consistent performance.