Row and column generation algorithms for minimum margin maximization of ranking problems

We consider the ranking problem of learning a ranking function from the data set of objects each of which is endowed with an attribute vector and a ranking label chosen from the ordered set of labels. We propose two different formulations: primal problem, primal problem with dual representation of normal vector, and then propose to apply the kernel technique to the latter formulation. We also propose algorithms based on the row and column generation in order to mitigate the computational burden due to the large number of objects.