Edge Features

[Roshan Santhosh]

At present, the joint feature vector that is created for a sample consists of only the counts of edges of a certain type. 

For example, in a binary classification problem, the unary features would simply be the sum of the node features for a particular class while the edge features would be the count of 0-0, 0-1 and 1-1 edges present in the entire graph. 

Is this really the best way to evaluate pair-wise potentials? Cant there be other ways that can better represent the pair-wise potentials and if so how can it be done in PyStruct?

Right now what I feel is that if say the weight for 0-0 edge is high. Then during inference it just tries to reduce the number of 0- classified labels so that the number of 0-0 edges come down.