Classification-Based Clustering Evaluation

The evaluation of clustering quality has proven to be a difficult task. While it is generally agreed that application-specific human assessment can provide a reasonable gold standard for clustering evaluation, the use of human assessors is not practical in many real situations. As a result, machine computable internal clustering quality measures (CQMs) are often used in the evaluation process. However, CQMs have their own drawbacks. Despite their extensive use in clustering research and applications, many CQMs have been shown to lack generality. In this paper we present a new CQM with general applicability. The basis of our CQM is a pattern recognition view of clustering's purpose: the unsupervised prediction of behavior from populations. This purpose translates naturally into our new classifier based CQM which we refer to as informativeness. We show that informativeness can satisfy core CQM axioms defined in prior research. Additionally, we provide experimental support, showing that informativeness can outperform many established CQMs by detecting a larger variety of meaningful structures across a range of synthetic datasets, while at the same time exhibiting good performance on each individual dataset. Our results indicate that informativeness provides a highly general and effective CQM.