About KGAT implementation

[Teresa0519]

Hello, I saw your question in pyG
I'm also currently working on KGAT using a pyG through Neo4j
Can I ask how you changed the structure of the original KGAT?

Thank you very much

[BWAAEEEK]

I tried to changed sampling method used for aggregation.
I mean when KGAT do propagation, sampled neighbor nodes is used for message passing.
So, to aggregate all of the neighbor nodes I was planning to implement all of the neighborhood nodes.

But, I have come to the realization that there is a problem when attempting to implement the propagation method in cases where two nodes have multiple edges of different types.
This presents a challenge for aggregating neighbor nodes, as using only one adjacency matrix may not be suitable.

If you look at the picture above, two nodes have two different types of edges.
And this one can not be handled by message passing method using one adjency matrix.

While PyG does offer HeteroConv as an option for aggregation, I am uncertain whether it is a suitable approach for the KGAT propagation method.
HeteroConv aggregates with multiple adjacency matrices independently, which may not align with the KGAT method.

I am also still grappling with implementing the model, and it would be great if we could share our experiences and any relevant information together, if you are also in the process of trying to implement something similar 😄