STAGE Representation and contrastive learning on graphs

In the literature, numerous graph neural network (GNN) models have been proposed for graph-related tasks, such as node classification, link prediction, and graph classification. Most existing GNN approaches use semi-supervised training. For many real-world graph applications, e.g. protein analysis, they intuitively require the input of some amount of labeled data.
Alternatively, there is a series of random walk-based GNNs, including node2vec and graph2vec,
which are unsupervised. Their approach is to first learn the node embeddings, and then various supervised downstream tasks are directly applied on these node embeddings. These approaches can be considered as part of the contrastive learning framework.

Contrastive learning originally aims to learn to embed each image in a self-supervised manner. Due to its impressive performance in many tasks, contrastive learning has become the hottest topic in unsupervised learning. Its motivation is to maximize the similarity of positive pairs and the distance of negative pairs. Generally speaking, the positive pairs are composed of data augmentations of the same instance, while those of different instances are regarded as negative pairs.

This internship aims to investigate the last findings in graph embedding and graph contrastive learning, understand, deconstruct the different steps, and make some steps towards new contrastive learning approaches for graphs.