Metaclustering and consensus algorithms for interactive data analysis and validation

Clustering of real-world datasets is a complex problem. Optimization models seeking to maximize a fitness function assume that the solution corresponding to the global optimum is the best clustering solution. Unfortunately, this is not always the case, mainly because of noise or intrinsic ambiguity in the data. In this work we present a set of tools implementing classical and novel techniques to approach clustering in a systematic way, with an application example to a complex biological dataset. The tools deal with the problem of generating multiple clustering solutions, performing cluster analysis on such clusterings (i.e. Meta Clustering) and reducing the final number of clusterings by the appropriate application of different Consensus techniques. A subsequent crossing of prior knowledge to the obtained clusters helps the user in better understanding its meaning and validates the solutions. © 2009 Springer Berlin Heidelberg.