How to tell when a clustering is (approximately) correct using convex relaxations

We introduce the Sublevel Set (SS) method, a generic method to obtain sufficient guarantees of near-optimality and uniqueness (up to small perturbations) for a clustering. This method can be instantiated for a variety of clustering loss functions for which convex relaxations exist. We demonstrate the applicability of this method by obtaining distribution free guarantees for K-means clustering on realistic data sets. Papers published at the Neural Information Processing Systems Conference.