Clustering via LP-based Stabilities
Nikos Komodakis, Nikos Paragios and Georgios Tziritas Poster ID: M23 Powerful clustering algorithm based on LP-duality theory:
Clusterings of almost optimal cost Online optimality bounds Automatic determination of number of clusters Independent of initialization (unlike, e.g., EM-like methods) Guaranteed convergence (no need to tweak any parameters) Applicable to any type of distances

Novel method for automatic cluster center selection:
LP-based stabilities LP-based margins newly introduced concepts dual to each other