Appendix for Diffusion Improves Graph Learning A Graph diffusion as a polynomial filter

L and use the binomial equation, i.e. To obtain a more convenient form for K we shift the summation index using m = k j, i.e. ξ All remaining nodes are part of the test set and only used once for testing. Different seeds are used for validation and test splits. The patience is reset after an increase in accuracy on the validation set. We use the same development set for optimizing the hyperparameters for clustering.