Diversity control for improving the analysis of consensus clustering

Consensus clustering has emerged as a powerful technique for obtaining better clustering results, where a set of data partitions (ensemble) are generated, which are then combined to obtain a consolidated solution (consensus partition) that outperforms all of the members of the input set. The diversity of ensemble partitions has been found to be a key aspect for obtaining good results, but the conclusions of previous studies are contradictory. Therefore, ensemble diversity analysis is currently an important issue because there are no methods for smoothly changing the diversity of an ensemble, which makes it very difficult to study the impact of ensemble diversity on consensus results. Indeed, ensembles with similar diversity can have very different properties, thereby producing a consensus function with unpredictable behavior. In this study, we propose a novel method for increasing and decreasing the diversity of data partitions in a smooth manner by adjusting a single parameter, thereby achieving fine-grained control of ensemble diversity. The results obtained using well-known data sets indicate that the proposed method is effective for controlling the dissimilarity among ensemble members to obtain a consensus function with smooth behavior. This method is important for facilitating the analysis of the impact of ensemble diversity in consensus clustering.