Why do we need add_self_loops?

[quy-ng]

Hi team, I knew there is a function named add_self_loops, but I struggle to understand why we should use it?

In the document, it is described Adds a self-loop to every node in the graph given by edge_index , but in test case occurs to me confusing.
edge_index = torch.tensor([[0, 1, 0], [1, 0, 0]])
expected = [[0, 1, 0, 0, 1], [1, 0, 0, 0, 1]]

Much appreciate if you could help me to understand

[wsad1]

So the add_self_loops function adds edges from each node to itself. Such a thing might be useful if you want to say pass a message from a node to itself , like lets say you have a graph with no self loops then a layer like GATConv will pay attention to only a nodes neighbors, but with add_self_loops it can pay some "attention" to itself too.

Now coming to the test case. add_self_loops doesn't check for existing self loops it simply appends all possible self loops to the edge_index so the expected output expected = [[0, 1, 0, 0, 1], [1, 0, 0, 0, 1]], the first three edges (0,1), (1,0), (0,0) are already present and the self loops (0,0), (1,1) are added the self loop (0,0) gets repeated.

On another note you can check add_remaining_self_loops which only adds self loops that are not already present.

[quy-ng]

thanks

[alexrutherford]

I wanted to ask if add_self_loops will solve the problem I have now. I basically want to train a GCN to fill in missing node level features. Currently, I have N node level features, and I want to be able to predict missing values in one of these features e.g. if a new node is added to the network (and I know which existing nodes it connects to).

So for training my features are the N node level features and the target is one of those N node level features. But obviously for the node I am interested in predicting, I only want to use the known feature values of the neighbours and to somehow mask the known feature value of that node of interest. Otherwise I am just predicting X based on X + {others}!

Would the deletion of self loops in the last layer effectively do that? I am looking for examples of predicting missing node level attributes but I couldn't find any.

[wsad1]

add_self_loops = False will prevent a node from having itself as a neighbour.
But for your problem, setting this to False will help for layers like GCNConv where the node update doesn't use target node information, but it won't help for layers like SAGEConv where node update depends on target node.

[rusty1s]

Please see my answer here #3428.