Exploring the Performance Limit of Cluster Ensemble Techniques

Cluster ensemble techniques are a means for boosting the clustering performance. However, many cluster ensemble methods are faced with high computational complexity. Indeed, the median partition methods are NP-complete. While a variety of approximative approaches for suboptimal solutions have been proposed in the literature, the performance evaluation is typically done by means of ground truth. In contrast this work explores the question how well the cluster ensemble methods perform in an absolute sense without ground truth, i.e. how they compare to the (unknown) optimal solution. We present a study of applying and extending a lower bound as an attempt to answer the question. In particular, we demonstrate the tightness of the lower bound, which indicates that there exists no more room for further improvement (for the particular data set at hand). The lower bound can thus be considered as a means of exploring the performance limit of cluster ensemble techniques.