Node Attributes and Edge Attributes about seal

[muhanzhang]

1. Yes. To use node class as attribute, you can prepare a group variable in your .mat graph, where each row in the group contains the one-hot-encoding of this node's class. You can refer to PPI_subgraph.mat for an example.

2. SEAL is easy to extend to multi-relational link prediction. You can use R-GCN as the GNN to handle multiple edge types, and replace the binary loss to multi-class loss. For example, see SEAL's sister model, IGMC, for recommender systems.

[LIANGKE23]

Thanks!
I have checked PPI-subgraph.mat, it is shown below.

There are 3889 rows and 49 collumns. In my understanding, it means there are 3889 nodes, and 49 kinds of attributes for all the nodes. For example, in the figure shown above, node 4 has two features which are feature 1 and feature 8. Am I right?

[muhanzhang]

Yes!