I want to cluster a set of different graphs ( the whole graphs). However, when I search for references from Google, I find there are few works related to it. So I want if there is anyone who has done this before.
I’m not aware of such work. One possibility is to first compute graph-level representations and then cluster them.
Thanks again, my friend!
So could please give me some hint about graph representation? (except graph2vec)
When you use GNN-based approaches for graph property prediction, the input to the MLP for prediction are graph-level representations.
But how can I design the model, a graph autoencoder or something?
What do you mean by design?
Just mean that how can I make an unsupervised model for representing.
There’s been an increasing interest in learning molecular representations via graph unsupervised/self-supervised learning. DGL-LifeSci allows computing molecular representations via some pre-trained models here.
Uhm, but I wonder if I can train a custom network myself since my data is kind of different with molecular data😂
Maybe you can check existing literature on unsupervised/self-supervised learning with GNNs and see if some of them are related to your scenario. If so, you can adapt the released code to your case.
Take a look at:
for an example. In this example, the resolution of a help desk service ticket is modeled as a graph. So the data is a collection of graphs. A heterogeneous graph convolution network is used to predict a node attribute. As a consequence, a graph embedding is also developed. HTH.