Modeling Edge-Rich Graphs Using Neural Networks

Abstract

The state of the art in graph analytics has advanced rapidly in recent years. Much of this advancement, to account for heterogeneous and dynamic networks, has been on graphs with rich node-level information such as those found on the Internet, social media, or in relational databases. Conversely, edge-rich graphs, where edges contain substantial information, have received relatively little attention. Examples of edge-rich graphs include networks derived from group communication logs and spatio-temporal movement modeling. In this paper, we use theoretical analysis and simulations to show that existing graph neural networks are less effective at handling such edge heterogeneity. We then propose EdgeCast, an edge-conditioned convolutional graph neural network approach that models edge-rich network data as context-enriched graphs for enhanced forecasting. We evaluate the proposed method against state-of-the-art feature-based and graph neural network methods on four real-world testbeds. Our results across simulations and real-world experiments suggest that emphasizing edge heterogeneity improves performance in tasks involving edge-rich graphs.