Clustering ensembles and space discretization – A new regard toward diversity and consensus

In recent years, the cluster ensembles have been successfully used to tackle well known drawbacks of individual clustering algorithms. Beyond the expected improvement provided by the averaging effect of many clustering algorithms (clustering committee) aiming at the same goal, some interesting experimental results also show that even committees of completely random partitions may lead to a useful consensus. Another powerful finding in cluster ensemble research is that the blind criterion Averaged Normalized Mutual Information seems to replace actual misclassification ratio, whenever labels are given to actual clusters. In this work, we study what is behind these interesting results and the blind criterion, and we use what we learn from this study to propose a new point of view for analysis and design of clustering committees. The usefulness of this new perspective is illustrated through experimental results.

Research highlights
► ANMI maximization is not a guarantee of good clusterings.
► Clustering Consensus can be regarded as a two-layered clustering task: first, in pattern space, and secondly, in label space.
► Clustering Consensus can be seen as a generalization of the Binary Morphology Clustering Algorithm (BMCA).
► Beyond the BMCA, Clustering Consensus may take advantage of the nonlinearity of the grid quantization provided by the clustering committee.