A finite time stochastic clustering algorithm

We present a finite time local search (1 + δ)-approximation method finding the optimal solution with probability almost one with respect to a general measure of within group-dissimilarity. The algorithm is based on a finite-time Markov model of the simulated annealing. A dynamic cooling schedule, allows the control of the convergence. The algorithm uses as measure of within group dissimilarity a new generalized Ward index based on a set of well-scattered representative points, which deals with the major weaknesses of partitioning algorithms regarding the hyperspherical shaped clusters and the noise. We compare it with other clustering algorithms, such as CLIQUE and DBSCAN.