Model Breaks down on using self loops

[SourabhRanade]

I have a model (consisting of a few GraphSage Layers followed by some torch.nn.Linear layers) that works perfectly fine when I train it on differently sized fully connected graphs without self loops. However, as soon as I add self loops to the graph using add_self_loops method and start the training, the model basically stops learning. I am not very sure why.
I also tried creating the self loops while I build my graphs (using networkx thus not using the add_self_loops method). I still end up with the model not learning anything. Each epoch basically gives me similar losses and accuracies.

Can someone guide me with this? Thank you in advance.

[rusty1s]

This is hard to tell. In the end, adding a self-loop as part of the graph means that your GNN can no longer differentiate neighbors from the central node, which may be a problem for your specific task/data.