search

WanyuGroup / CVPR2022-OrphicX

Official code for the CVPR 2022 (oral) paper "OrphicX: A Causality-Inspired Latent Variable Model for Interpreting Graph Neural Networks."
MIT License
34 stars 14 forks source link

graph datasets #2

Open Usama113 opened 2 years ago

Usama113 commented 2 years ago

Hi, I could not find the graph classification datasets Mutagenicity and NCI1 in the repo. Could you please point me. Also, when loaded from the ckpt folder, each graph of Mutagenicity dataset have [100,100] dense representation of Adj matrix. According to my findings, some graphs in Mutagenicity dataset have variable number of nodes in a graph some exceeding 100 nodes per graph. In that case the Adj matrix will not [100,100]. Could you please also share how the dense representation of Adj was created. Thanks in advance.

haolanut commented 2 years ago

Hi,

GNNExplainer use a fixed graph size of 100 nodes and skips all graphs that have more than 100 nodes. We use GNNExplainer's code to process the dataset and train the graph classification model. Then, we save the processed dataset and graph classification model parameters into the checkpoint. In orphicx, we directly load the checkpoint and then train our explainer.