questions about the classifier in mmdcgm

[ifangcheng]

In the mmdcgm, the prediction of a classifier yn=argmax( q(y|xn) ) is used as a point estimation to approximate the full posterior q(y|xn).

The question is that during the training process, what exactly the classifier is ?
Is it defined by yn=argmax( q(y|xn) ) or by a max-margin classifier: y=argmax F(y;x),
what's the relation between yn=argmax( q(y|xn) ) and a max-margin classifier: y=argmax F(y;x) ?
and when testing, which classifier to adopt?