Graph neural networks (GNNs) are a topic of increasing interest in recent years. They have the potential to handle irregularly structured data, which is of great interest for graph-structured data. Like other areas, graph-structured data suffers from a lack of data. The available data consists of only a small portion of labeled data, while the majority is unlabeled. Semi-supervised learning with contrastive learning has been successfully applied in image representation learning to solve this problem. Contrastive learning relies on good augmentations, which is more complicated for graph-structured data and still needs further exploration. In this thesis, we attempt to solve the problem by creating multiple new augmentations and comparing them to existing ones. We evaluate the performance of these single augmentations and test the combination of augmentations with different augmentation ratios. Additionally, we further improve the results by testing different loss functions. Eventually, we test transfer learning and warm-starting the neural network with the same and different datasets. Ultimately, we give an outlook into improved ideas to individually select augmentations for each graph and connect them to recent research that uses generators to create augmentations.