Graph Neural Networks for Semi-Supervised Learning on Social Networks

Abstract

Graph Neural Networks (GNNs) have emerged as powerful models for learning on graph-
structured data, particularly for tasks like node classification, link prediction, and graph
clustering. This paper focuses on the application of GNNs to semi-supervised node
classification in social networks, where labeled data is scarce but the network structure is rich.
We implement and evaluate Graph Convolutional Networks (GCN), Graph Attention Networks
(GAT), and GraphSAGE on three public benchmarks: Cora, Citeseer, and PubMed. GATs
outperform other models in accuracy (up to 84.7% on Cora) due to their ability to assign
learnable importance weights to neighboring nodes. GraphSAGE demonstrates scalability
advantages for large graphs through neighborhood sampling. We explore the effect of feature
normalization, activation functions, and layer depth on classification accuracy and training
stability. Our experiments confirm that two-layer GNN architectures strike the best balance
between expressiveness and overfitting risk. We also assess model robustness to adversarial
perturbations and missing features. The results validate GNNs as state-of-the-art methods for
semi-supervised learning on social network data, enabling accurate label propagation through
graph topology. This paper contributes a practical guide to selecting and tuning GNN models
for network-centric machine learning applications.