Mis on kontrastne õppimine graafikutes (DGI, GraphCL)?

Contrastive learning on graphs is a method of learning representations of nodes or graphs based on comparing positive and negative pairs of examples to identify significant features.

Specifically:

• DGI (Deep Graph Infomax) trains the model by maximizing mutual information between local (node) and global (graph) representations. This is achieved by contrasting real graphs with negatives (e.g., noisy or random versions).

• GraphCL uses various graph augmentations (such as edge removal, node masking) to create positive pairs, while other graphs or augmentations are considered negatives. The model learns to distinguish positive pairs from negatives, improving the quality of embeddings.

In this way, contrastive learning on graphs helps to train robust and informative representations without the need for large amounts of labeled data.