snap-research / mlpinit-for-gnns Goto Github PK

Hi! thanks for providing such an insightful paper! I'm curious about whether the MLPInit can deal with data with missing node features, which are common in practice. In this case, MLP may not be well trained since graph structure is unavailable. So how your method be applied in such case?

Thanks in advance for your help!

Hey! I came across your paper MLPinit and I am trying to implement the PEERMLP for the GAP paper: https://arxiv.org/abs/2203.00949

I aim to apply transfer learning on GAP to achieve the same privacy budget in less number of epochs.

The proposed MLPInit method does not even need extra data for pre-training. I can simply use the node features to train the PeerMLP of GAP and then use its learned weights to initialize GAP. Under edge-level DP, this approach does not incur any privacy costs.

Could you please guide me on How can I make changes in the GAP code and train its PEERMLP?
GAP repo: https://github.com/sisaman/GAP