Generating Attention Weights From GATCONV

[akul-goyal]

Hi, I am trying to understand the embedding of certain nodes within my graph. I am using a 2 layers Graph Attention Neural Network to create the embeddings. I want to find out the weights that are placed for each of the nodes within a k-hops of a target node. Is there a way for me to do this?

[rusty1s]

You can get the attention-based adjacency matrix from GATConv by making use of return_attention_weights=True in its forward call. Together with utils.k_hop_subgraph(), you should be able to gather the weights of a node within its k-hop neighborhood.

[akul-goyal]

Thanks for the quick response. I will give it a try!