DebiasingGraphNeuralNetworksviaLearning DisentangledCausalSubstructure

With the disentangled representations, we synthesize the counterfactual unbiased training samples to further decorrelate causal and bias variables.