A stochastic approach for clustering in object bases

Object clustering has long been recognized as important to the performance of object bases, but in most work to date, it is not clear exactly what is being optimized or how optimal are the solutions obtained. We give a rigorous treatment of a fundamental problem in clustering: given an object base and a probabilistic description of the expected access patterns, what is an optimal obJect clustering, and how can this optimal clustering be found or approximated? We present a system model for the clustering problem and discuss two models for access patterns in the system. For the first, exact optimal clustering strategies can be found; for the second, we show that the problem is NP-complete, but that it is an instance of a well-studied graph partitioning problem. We propose a new clustering algorithm based upon Kernighan’s heuristic graph partitioning algorithm, and present a preliminary experimental comparison of this new clustering algorithm with several previously proposed clustering algorithms.