Joint data augmentations for automated graph contrastive learning and forecasting

Abstract Graph augmentation plays a crucial role in graph contrastive learning. However, existing methods primarily optimize augmentations specific to particular datasets, which limits their robustness and generalization capabilities. To overcome these limitations, many studies have explored automated graph data augmentations. However, these approaches face challenges due to weak labels and data incompleteness. To tackle these challenges, we propose an innovative framework called Joint Data Augmentations for Automated Graph Contrastive Learning (JDAGCL). The proposed model first integrates two augmenters: a feature-level augmenter and an edge-level augmenter. The two augmenters learn whether to drop an edge or node to obtain optimized graph structures and enrich the information available for modeling and forecasting task. Moreover, we introduce two stage training strategy to further process the features extracted by the encoder and enhance their effectiveness for forecasting downstream task. The experimental results demonstrate that our proposed model JDAGCL achieves state-of-the-art performance compared to the latest baseline methods, with an average improvement of 14% in forecasting accuracy across multiple benchmark datasets.