Learning invariant representation on non-homophilous graphs is crucial to address distribution shifts and biases.
