Graph Neural Networks for outlier detection

Abstract

The thesis extensively explored state-of-the-art Graph Neural Networks (GNNs) for nodelevel outlier detection within graph data. A comprehensive review of various GNN architectures and outlier detection algorithms was conducted. Using PyTorch and the PyGOD library, the performance of four node-level outlier detection algorithms, DOMINANT, AnomalyDAE, CoLA, and GAAN was evaluated on the Cora and CiteSeer datasets, which were manually injected with 50 node-level outliers.

The models were assessed based on their AUC scores derived from ROC curves. AnomalyDAE and DOMINANT exhibited the highest performance, achieving AUC scores of ~0.81 and ~0.83 for the Cora dataset, and ~0.80 and ~0.83 for the CiteSeer dataset, respectively. CoLA followed closely with AUC scores of ~0.78 for Cora and ~0.80 for CiteSeer while GAAN demonstrated comparatively lower performance, with AUC scores of ~0.74 for Cora and ~0.78 for CiteSeer. Detection in node-level outliers where only 100 features were altered presented challenges across models, with variations observed in AUC scores. However, all models identified every node-level outliers where every features were altered.