Addressing overfitting in graph convolutional network using data augmentation

Abstract

This project investigates the application of Graph Convolutional Networks (GCNs) to functional brain networks for subject classification, with a focus on addressing model overfitting through hyperparameter optimization and data augmentation. Experiments were conducted on five publicly available neuroimaging datasets: TaoWu, Neurocon, PPMI, ABIDE, and ADNI using fully connected graphs derived from fMRI connectivity matrices. The study proceeded in two phases. The first phase utilized unweighted graphs to establish a baseline and evaluate traditional regularization techniques, including dropout and DropEdge. The second phase incorporated weighted graphs to better capture connectivity strength, and introduced two edge-aware augmentation strategies: Gaussian edge weight perturbation and edge weight scaling. Hyperparameter optimization consistently improved test accuracy and reduced performance variability across datasets. However, dropout and DropEdge failed to mitigate overfitting, and edge scaling yielded diminishing results across all datasets. Gaussian perturbation led to modest improvements in larger datasets but negatively affected smaller ones. Performance remained below that of domain-informed, semisupervised models from prior work, underscoring the limitations of imaging-only supervised learning. Across both stages, hyperparameter tuning was the most reliable source of performance gains. Moreover, persistent overfitting and plateauing validation performance pointed to limitations in the augmentation strategies and training dynamics. The findings highlight the importance of tuning, the sensitivity of GCNs to augmentation design, and the need for domain-aware modelling. Recommendations are made for future work, including the integration of multimodal data, augmentation parameter tuning, non-stochastic augmentation techniques, and improving early stopping strategies.