Search for statistically significant clusters in networks and biomedical applications

Abstract

A new, statistically based method of clustering previously used by authors for analysis of frequency and point data in epidemiology and for searching clusters in samples is applied to discovering clusters in graphs or networks. Graph clustering or community detection is often used in biomedical applications. Nevertheless methods existing in literature have no sufficient statistical background and the question of significance for revealed clusters usually is not considered. In the case when clusters do not form the whole network false positive results are inevitable. We propose method for discovering statistically significant clusters. It is compared with nine methods existing in literature. It is shown that only proposed method can discover significant clusters successfully discriminating them from phone consisting from randomly connected network objects.