Unsupervised stability-based ensembles to discover reliable structures in complex bio-molecular data

The assessment of the reliability of clusters discovered in bio-molecular data is a central issue in several bioinformatics problems. Several methods based on the concept of stability have been proposed to estimate the reliability of each individual cluster as well as the ”optimal” number of clusters. In this conceptual framework a clustering ensemble is obtained through bootstrapping techniques, noise injection into the data or random projections into lower dimensional subspaces. A measure of the reliability of a given clustering is obtained through specific stability/reliability scores based on the similarity of the clusterings composing the ensemble. Classical stability-based methods do not provide an assessment of the statistical significance of the clustering solutions and are not able to directly detect multiple structures (e.g. hierarchical structures) simultaneously present in the data. Statistical approaches based on the chi-square distribution and on the Bernstein inequality, show that stability-based methods can be successfully applied to the statistical assessment of the reliability of clusters, and to discover multiple structures underlying complex bio-molecular data. In this paper we provide an overview of stability based methods, focusing on stability indices and statistical tests that we recently proposed in the context of the analysis of gene expression data.