Investigating Properties of Contrastive Learning in Geometric Deep Learning
While deep learning methods have achieved tremendous advancements within the last decade, it only works with Euclidean data, which clearly does not support non-Euclidean data types such as graphs. Geometric deep learning tries to employ and generalize deep learning methods to graphs. However, labeling graphs due to their domain-specific and complex nature is not always straightforward, resulting in label scarcity for graphs that necessitates employing self-supervised methods such as contrastive learning in geometric deep learning. This project aims to investigate different properties and effects of contrastive learning in Graph Neural Networks.