Graph Neural Networks in Complex Data Modeling

Abstract

Graph neural networks extend deep learning to relational data by operating directly on graph-structured inputs - molecular bonds, social connections, citation links, traffic networks -- where the topology itself carries predictive information that flat feature vectors discard. This study presents a controlled empirical evaluation of six GNN architectures -- GCN, GAT, GraphSAGE, GIN, PNA, and GPS (General Powerful Scalable) -- across five complex data modelling tasks: molecular property prediction (OGB-MolHIV and QM9), social network community detection (Reddit, Flickr), citation network node classification (ogbn-arxiv, Cora), traffic flow forecasting (METR-LA, PEMS-BAY), and protein function prediction (PPI). A total of 2,160 experiments were conducted under standardised hyperparameter budgets and hardware. GPS achieved the highest mean performance across tasks (mean rank 1.8 across 10 datasets), followed by PNA (2.4) and GAT (3.2). GIN achieved the best results on graph-level classification tasks (OGB-MolHIV AUC-ROC = 0.802 +- 0.012) where distinguishing non-isomorphic graphs is critical. GAT achieved the strongest node classification results (ogbn-arxiv accuracy = 73.8 +- 0.4%) through learned attention over neighbour importance. Over-smoothing limited all architectures to 4-6 effective layers on homophilic graphs, with GPS's positional encoding and global attention partially mitigating this to 8-10 layers. Scalability analysis revealed that mini-batch sampling (GraphSAGE-style) was essential for graphs exceeding 100K nodes, where full-batch methods exceeded GPU memory. A practical architecture selection framework mapping graph characteristics to recommended GNN configurations is proposed.