1764183ef03fc7324eb58c3842bd9a57-Supplemental.pdf

A.1 Datasets Table S1 reports summary statistics of the datasets used in this paper. For SMNIST and ZINC, we use the same pre-processing steps and data splits as in [10]. Errica et al.[11] show that drawing conclusions based on some of these datasets can be problematic as structure-agnostic baselines achievehigher performance than traditional GNNs. However,intheir assessment, NCI1 is the only chemical dataset on which GNNs beat baselines. A.2 Models We implement all models using the PyTorch Geometric Library [12].