In real life, heterogeneous graphs have different entities with different features. The HAN should utilize heterogeneous nodes’ features as the author said. However, in our code, we only use features about the “paper”. What if I want to add more features about different entities such as the author? How to change the code?
I don’t think HAN’s formulation allows using features of node types different from what you are predicting (e.g. they should always be papers, not authors etc.). The reason is that the metapaths must start and end at the same node type.
I think you are right. Once the meta path-based neighbors are created, we only have one type of node.
However, what if we change the metapaths? “A-P-A” only learns features from the Author, we change it into PAP, and then we aggregate them.
That makes sense. You could have different metapaths for each node type and learn representations for each of them.
This topic was automatically closed 30 days after the last reply. New replies are no longer allowed.