Input Features for Chameleon and Squirrel

[YiteWang]

Thank you for the amazing work! I am new to GNN and I am a little confused why you change the input feature of Chameleon and Squirrel to one hot encoding in https://github.com/LUMIA-Group/OrderedGNN/blob/main/task_node.py#L23:

data.x = torch.eye(data.x.shape[0])

Thank you so much!

[realCrush]

Thank you for your interest! We use the setting described in Section 4.3 of [1] for Chameleon and Squirrel. In practice, we have found that the node features in these two datasets are less informative and can potentially mislead our gating module, which is purely data-driven. Therefore, we replace the input feature with one-hot encoding so that GNN mainly relies on the graph structure. You can refer to Table 4 and 5 in [1] to check the performance of the baseline models under this setting.

[1] Graph Neural Networks with Heterophily, AAAI 2021

[YiteWang]

Thank you for the explanation!