Pattern Recognition, 4th Edition

16.4. Relative Criteria

So far, clustering validation has been performed on the basis of statistical tests. A major drawback of most of these techniques is their high computational demands, due to the required Monte Carlo methodology. In this section, a different approach is discussed that does not involve statistical tests. To this end, a set of clusterings is considered and the goal is to choose the best one according to a prespecified criterion. More specifically, let A be the set of parameters associated with a specific algorithm. For example, for the algorithms of Chapter 14, A contains the number of clusters, m, as well as the initial estimates of the parameter vectors associated with each cluster. The problem can be stated as follows: ...