Learning from Graphs Beyond Message Passing Neural Networks

Tong Zhao, Neil Shah, Elham Ghazizadeh

Keywords: Graph Machine Learning

Abstract: Graph, as a potent data structure, models complex relational data that are ubiquitous in real-world applications like social networks and recommendation systems. In the past few years, message passing-based Graph Neural Networks (GNNs) have emerged as standard tools for direct learning from graph data. However, such direct integration during training also introduces challenges, including scalability issues with large-scale graphs and oversmoothing problems with increased model parameter sizes via additional layers. This study offers a bird's-eye view of high-level paradigms for learning from graph data, categorizing techniques into three distinct classes: (1) using graph structure during preprocessing, (2) using graph structure during training, and (3) using graph structure at test-time inference. Through this overview, we aim to illuminate diverse approaches and advantages inherent in learning from graph data across these fundamental paradigms.

Submission Number: 173