GNN efficiency evaluation for node classification across multi-domain datasets with different noise conditions

Node classification in graph data is critical for selecting the optimal algorithm. In this study, we analyze different datasets with various Graph Neural Network (GNN) models to evaluate node classification accuracy. The primary goal is to identify efficient methods for multi-domain datasets that can outperform other approaches. We selected 14 different datasets from various domains for our experiment. To find an effective solution for node classification, we explored 10 GNN models for accuracy comparison. Each model was tested with different settings, and the best-performing model was selected. We compared the node classification accuracy for each model. Next, we adopted a resilience-checking approach to find the best models that perform well under different noise conditions. We employed three different strategies: Gaussian, label, and edge-level noise to assess the ability of different models in node classification. The experimental results show that some models demonstrate superior capacity in node classification, including SAGEConv, ChebConv, and H2GCN. SAGEConv shows superior resilience against noise whereas ChebConv has good node classification accuracy.