Post #86

But then, they say that GIN in fact uses MLP as the aggregation function \phi and they only hope that it will learn injective function. There is no constraint on the MLP so in principle it can easily learn non-injective function, in which case it will be less powerful then 1-WL.