A clustering algorithm for data-sets with a large number of classes

摘要

A goal of a clustering algorithm is to minimise a cost metric, but existing methods have a tendency to reach a local minimum of the metric. This paper introduces a new clustering algorithm which can jump out of a local minimum for a clustering problem with a large number of classes. This algorithm is developed from an interactive one which employs four operations to change the cluster distribution; knowledge about how to get a lower local minimum of the metric has been implanted into the new algorithm. Without interactive operation the new algorithm can try more changes to a cluster distribution, thereby enhancing its chance of reaching a better result. Because a machine can select an operation much faster than a human operator the new algorithm also needs less time to reach the final clustering result. All these points have been demonstrated by experiment.