Approximation algorithms for classification problems with pairwise relationships: metric labeling and Markov random fields
In a traditional classification problem, we wish to assign one of k labels (or classes) to each of n objects, in a way that is consistent with some observed data that we have ab...