Validating Explorative Clustering on a Hold-Out Test Set

Comparing clustering algorithms is much more difficult than comparing classification algorithms, which is due to the unsupervised nature of the task and the lack of a precisely stated objective. We consider explorative cluster analysis as a predictive task (predict regions where data lumps together) and propose a measure to evaluate the performance on an hold-out test set. The performance is discussed for typical situations and results on artificial and real world datasets are presented for partitional, hierarchical, and density-based clustering algorithms. The proposed S-measure successfully senses the individual strengths and weaknesses of each algorithm.