Supervised GNNs for Node Label Classification in Highly Sparse Network: Comparative Analysis

Graph Neural Networks (GNNs) have emerged as a powerful tool in machine learning for the modeling, analysis, and prediction of complex interaction networks. Our research is centered around a comparative analysis of three distinct GNN architectures: GCN, GAT, and GraphSAGE. Through this investigation, our goal is to provide insights into the effectiveness of GCN, GAT, and GraphSAGE in addressing the challenges associated with node classification within extremely sparse networks. We applied our methodology to a Twitter dataset, enriched with ground truth labels categorizing users as misinformation spreaders or non-spreaders in nine categories. Our focus is on exploring GNNs for highly sparse networks, aiming to identify effective methods for challenges in very sparse graphs.Our observations revealed that GCN performed exceptionally well with high accuracy across various categories, outshining GAT and GraphSAGE. This suggests that GCN may be superior in classifying nodes, especially in situations with limited data and a small number of links for a given number of nodes. Surprisingly, GAT showed better accuracy and precision than GraphSAGE, challenging the assumption that even they vary depending on the specific task and dataset, but generally GraphSAGE is better on accuracy and precision compared to GAT for smaller networks. This unexpected result prompts a deeper exploration of Graph Neural Network architectures and emphasizes the importance of aligning theoretical assumptions with real-world outcomes, particularly in sparse network scenarios. Our contribution involves a new perspective on different GNN architectures for highly sparse networks.