Firstly, SGTs ( i.e., simple GNNs) are still powerful and can "distinguish almost all non-isomorphic graphs" [ VQ-V AE (Table 3b) emphasizes the impact of pretraining methods on the tokenizer's performance. We leave the investigation of how to effectively pretrain GNN-based tokenizers as future works. We have included the literature review of MGM in the main body of the paper. However, a closer inspection reveals several critical distinctions between MGM and these methods. Finally, MGM employs remask decoding to constrain the encoder's ability on This code uses a single-layer SGT of GIN as an example.

In particular the generative adversarial networks (GANs) [3] provide an impressively powerful waytorepresent classes ofsignals.

Wetheoretically showour method converges for smooth objectives with square regularizers and the convergence dependence on the projection dimension is mild. We also illustrate the benefits of robustness and fairness on a class of linear problems.