Node classification on HeteroData() (Heterogeneous graph)

[maryamhgf]

Hi,

I have a dataset which is representing the computational graph of DNN models. I am using HeteroData().
Data:
Nodes: operators (Conv, Add, etc.)
Edges: the relationship between two operators ('Constant', 'connects', 'Add')
An overview of the data structure:
HeteroData( Constant={ x=[301, 8] }, Add={ x=[211, 5] }, Mul={ x=[203, 5] }, Conv={ x=[156, 20] }, Transpose={ x=[125, 10] }, ReduceMean={ x=[125, 7] }, Div={ x=[114, 5] }, Sub={ x=[64, 5] }, Pow={ x=[65, 5] }, Sqrt={ x=[66, 5] }, Erf={ x=[53, 5] }, GlobalAveragePool={ x=[3, 5] }, Flatten={ x=[3, 6] }, Gemm={ x=[3, 8] }, (Constant, connects, Constant)={ edge_index=[2, 283] }, (Constant, connects, Add)={ edge_index=[2, 23] }, (Constant, connects, Conv)={ edge_index=[2, 1] }, (Conv, connects, Mul)={ edge_index=[2, 21] }, (Mul, connects, Add)={ edge_index=[2, 23] }, (Add, connects, Conv)={ edge_index=[2, 9] }, (Add, connects, Add)={ edge_index=[2, 9] }, (Conv, connects, Transpose)={ edge_index=[2, 10] }, (Transpose, connects, ReduceMean)={ edge_index=[2, 12] }, (Transpose, connects, Sub)={ edge_index=[2, 12] }, ... (Conv, connects, Add)={ edge_index=[2, 1] }, (Transpose, connects, GlobalAveragePool)={ edge_index=[2, 1] }, (GlobalAveragePool, connects, Flatten)={ edge_index=[2, 1] }, (Flatten, connects, Gemm)={ edge_index=[2, 1] } )

As can be seen, the features of each node type are different from the other ones. I want to add a node and based on its connections, predict its type (node classification). Here type is the operator type (Add, Conv, etc.).

I am not sure which network structure I should use. I appreciate any suggestions, thoughts, etc. I have tried HeteroGNN. But, since the features are different for each node type, I am not sure how I can implement it correctly.

[wsad1]

What do you mean by But, since the features are different for each node type. Do you mean the feature dimension is different for each node type? If so heterogenous GNNs should be able to handle node types with different feature dimension.
Check out this example.