I’m using GCN for a graph classification task. In my case, the structures of graphs are very similar. The main differences between graphs are the node features. The dimension of the node feature is 1024. In one graph, the numbers of nodes are between several hundred and several thousand. And edges are on the tens of thousands level. I started with 2 layers of gcn and the feature dimension is 1024.
What I have tried:
adding more gcn layers, decrease the gcn feature dimension, using dgl.sum_nodes as readout all of them does not help to improve the training. My highest training accuracy is 60% and validation accuracy is under 10%.
What’s in my mind now is I’m having an underfitting problem and I will probably neet to increase the size of my network. But I’m relatively new to graph deep learning, so I’m wondering maybe there are some tricks already existing to deal with this kind of problem in graph deep learning? Any comments are appreciated. Thanks