A needle in a haystack

摘要: This paper addresses the problem of finding a small and coherent subset points in given data. problem, sometimes referred to as one-class or set covering, requires find small-radius ball that covers many data possible. It rises naturally wide range applications, from gene-modules extracting documents' topics, where are irrelevant task at hand, applications only positive examples available. Most previous approaches this focus on identifying discarding possible outliers. In we adopt an opposite approach which directly aims coherently structured regions, by using loss function focuses local properties We formalize learning optimization Information-Bottleneck principle. An algorithm solve is then derived analyzed. Experiments gene expression text document corpus demonstrate merits our approach.