Abstract  Cluster ensembles generate a large number of different clustering solutions and combine them into a  more robust and accurate consensus clustering. On  forming the ensembles, the literature has suggested  that higher diversity among ensemble members  produces higher performance gain. In contrast,  some studies also indicated that medium diversity  leads to the best performing ensembles. Such contradicting observations suggest that different data,  with varying characteristics, may require different  treatments. We empirically investigate this issue by  examining the behavior of cluster ensembles on  benchmark data sets. This leads to a novel framework that selects ensemble members for each data  set based on its own characteristics. Our framework  first generates a diverse set of solutions and combines them into a consensus partition P*. Based on  the diversity between the ensemble members and  P*, a subset of ensemble members is selected and  combined to obtain the final output. We evaluate  the proposed method on benchmark data sets and  the results show that the proposed method can significantly improve the clustering performance, often  by a substantial margin. In some cases, we were  able to produce final solutions that significantly  outperform even the best ensemble members