We use GIN Xu et al. (2018a) while DGI uses GCN Kipf & Welling (2016) as GIN provides a better inductive bias for graph level applications

For example, we use sum over mean for READOUT and that can provide important information regarding the size of the graph.

However, supervised tasks and unsupervised tasks may favor different information or a different semantic space. Simply combining the two loss functions using the same encoder may lead to “negative transfer”

Last update : February 13, 2023
Created : February 13, 2023