Fig. 4.

The general framework of contrastive methods. A contrastive method can be determined by defining its views generation, encoders, and objective. Different views can be generated by a single or a combination of instantiations of three types of transformations. Commonly employed transformations include node attribute masking as feature transformation, edge perturbation and diffusion as structure transformations, and uniform sampling, ego-nets sampling, and random walk sampling as sample-based transformations. Note that we consider a node representation in node-graph contrast [13, 38, 41] as a graph view with ego-nets sampling followed by a node-level encoder. For graph encoders, most methods employ graph-level encoders and node-level encoders are usually used in node-graph contrast. Common contrastive objectives include Donsker-Varadhan representation, Jensen-Shannon estimator, InfoNCE, and other non-bound objectives. An estimator is parametric if projection heads are employed, and is non-parametric otherwise. Examples of three specific contrastive learning methods are illustrated in the figure. Red lines connect options used in MVGRL [10]; green lines connect options adopted in GraphCL [49]; yellow lines connect options taken in GCC [48].