In cluster ensembles, we reuse only previous labelings and do not allow access to the original features. In many application, at least limited information from the original features is available at the combination stage. Primarily this is the case when the set of clusterings is generated by intentional splitting of the data.
Using extra feature information might further boost results and allow the cluster ensemble to be applied in a wider class of knowledge reuse applications. For example, in distributed clustering the minimum information for consolidation are the individual labelings. Adding extra feature information such as cluster centroids to augment the cluster ensemble can possibly improve results significantly.