Stability-based cluster analysis applied to microarray data

This paper studies the estimation of the number of clusters using the so-called stability-based approach, where clusters obtained for two subsets of the dataset are compared via a similarity index and the decision regarding the number of clusters is taken based on the statistics of the index over randomly selected subsets. We introduce a new similarity index s(., .), and analyze the consistency of the estimator of the number of classes when k-means algorithm is used in conjunction with s(., .). Various similarity indices are experimentally evaluated when comparing the "true" data partition with the partition obtained at each level of a hierarchical clustering tree. Finally, experimental results with real data are reported for a glioma microarray dataset.