ANGraph: attribute-interactive neighborhood-aggregative graph representation learning

We study the graph representation learning problem that has emerged with the advent of numerous graph analysis tasks in the recent past. The task of representation learning from graphs of heterogeneous object attributes and complex topological structures is important yet challenging in practice. We propose an Attribute-interactive Neighborhood-aggregative Graph learning scheme (ANGraph), which better preserves structure proximity and attribute affinity by leveraging attribute interaction and smoothness measures to incorporate vertices similar/close to each other in the original space. In addition, the proposed neighborhood aggregation mechanism aims to improve the performance of GNNs in various types of graphs. Evaluations carried out on five node classification datasets with different information densities verify the benefits of discriminative utilization of vertex information and show that ANGraph significantly outperforms the state-of-the-art methods.

Automated summary

The study focuses on graph representation learning, proposing a new learning scheme (ANGraph) that better preserves the characteristics of graph structures and achieves significant performance improvement in node classification tasks.