Simplifying-Clustering-with-Graph-Neural-Networks
Simplifying-Clustering-with-Graph-Neural-Networks copied to clipboard
FilippoMB
How to use for graph level unsupervised clustering, not the node?
I have millions of graph data that cannot be labeled by human efforts. I need it to be classified by unsupervised technique, not semi and not the node level. Is it possible?
No, sorry, not with this approach. You need something else.
The first step would be to generate some graph-level embedding. You can do that, for example, by using some message-passing layer to generate node embedding vectors and then aggregate them all together with some global pooling operation (e.g., you sum all the node embedding vectors).
Now the tricky part: you need an unsupervised loss to train your model to produce embeddings that are actually meaningful. You can look at something like this: https://arxiv.org/abs/1904.01098.
The paper I linked uses a graph edit distance, which has several drawbacks in my opinion. Probably there are more recent methods that work better and are simpler to use. Let me know if you find something interesting :)
