Closed lihuiliullh closed 3 years ago
The code can't do subgraph matching out of the box. Currently it's set up for subgraph classification. It assumes that you have prespecified subgraphs & associated labels that you want to predict.
However, the SubGNN architecture is generalizable and can be used for unsupervised or semi-supervised learning as well by changing the loss function. I can imagine using an approach like that used in NeuroMatch to identify query subgraphs and then using SubGNN with a semi-supervised learning objective to embed the subgraphs.
Hope this helps!
A very nice paper. But can this method be used to do subgraph matching?