Rethinking pooling in graph neural networks -- Supplementary material -- A Implementation details

Table S1 reports summary statistics of the datasets used in this paper. Table S1 also shows statistics for PROTEINS, NCI109, DD, and MOLHIV . For all datasets, we employ an initial convolution to extract node embeddings. The Complement model sticks to exactly the same setup. We use mini-batches of size 64 for SMNIST and ZINC.