SLAPS: Self-Supervision Improves Structure Learning for Graph Neural Networks - Supplementary Material

In the main text, we described how some edges may receive no supervision during latent graph learning. Specifically, we tested SLAPS without self-supervision (i.e. The accuracies for 2, 4, and 6-layer models are 66.2%, 67.1%, and 55.8% respectively. It can be viewed that increasing the number of layers from 2 to 4 provides an improvement. Work was done when authors were at Borealis AI. 35th Conference on Neural Information Processing Systems (NeurIPS 2021).