The dimension of H(t) and transform matrix W and U in CGNN/ODE

[MaiSpace]

Hello, thank you for sharing such a nice work !
I have some questions about the dimension of the H(t) and transform matrix W and U in STGODE. I wonder whether the following statements are right ?

• As you mentioned in the paper, classic form of GCN (H‘=AHW) entangles feature channels using a matrx W with dims [C, C'].
  However, when using CGNN with matrix-based or tensor-based ODE, the dims[0] and dims[1] of feature transform matrix W and temporal transform matrix U should be the same (dims: [F, F] or [T, T]) to make sure that H(t) has the same shape with H0 ?
• And to keep the dims of feature channels and temporal channels unchanged, the model fixed the feature dims at F in each layer, and pad zeros after dilated TCN to keep sequence length at T ?

Can STGODE model deal with the conditions with variable num of feature channels ?

Looking forward to your reply and thank you very much !