Graph Neural Networks: Bridging Structure and Intelligence

Abstract

Graph Neural Networks (GNNs) have emerged as a powerful paradigm for learning on non-Euclidean data, bridging the gap between structured graph representations and intel ligent data processing. This thesis explores the theoretical foundations and architectural innovations of GNNs, focusing on their ability to capture complex relational patterns in graph-structured data. We begin by examining the Message Passing Framework, which forms the backbone of most GNN architectures, enabling information propagation across graph nodes. The study then delves into Graph Convolution Operations, highlighting how they generalize traditional convolutions to irregular graph domains. We provide an in-depth analysis of two key GNN architectures: Graph Convolutional Networks (GCNs) and Graph Attention Networks (GATs). GCNs are presented as a foundational model that efficiently aggregates neighborhood information, while GATs are explored for their adaptive atten tion mechanisms that allow for more flexible and expressive feature learning. Through a comparative analysis, we elucidate the strengths and limitations of these architectures, with particular emphasis on citation network analysis, where nodes represent scientific papers and edges represent citations between them. Our experiments on the CORA dataset, com prising 2,708 scientific publications and 5,429 citation links, demonstrate the effectiveness of GNNs in capturing the interdependence between academic papers and their citations.