MESSAGE PASSING GRAPH NEURAL NETWORKS

Abstract

Graph Neural Networks (GNNs) have emerged as powerful tools for learning representations of graph structured data, finding applications in diverse fields such as social network analysis, recommendation systems, and molecular chemistry. In this master’s thesis, we present a thorough examination of state-of-the-art message-passing neural networks, including but not limited to Graph Convolutional Networks (GCN), Graph Attention Networks (GAT), GraphSAGE, and Approximate Personalized Propagation of Neural Predictions (APPNP). Our study begins with a detailed exploration of the foundational concepts and theoretical underpinnings of message-passing neural networks. We delve into the fundamental mechanisms of information propagation across nodes within a graph, elucidating the various aggregation and diffusion strategies employed by different architectures. Subsequently, a comprehensive review of existing GNN models is undertaken, providing a comparative analysis of their strengths, weaknesses, and unique characteristics. This work contributes a concise yet comprehensive overview, aiding researchers and practitioners in understanding the evolving landscape of graph representation learning.