Semi-possibilistic Biclustering Applied to Discrete and Continuous Data

Abstract

In contrast to hard biclustering, possibilistic biclustering not only has the ability to cluster a group of genes together with a group of conditions as hard biclustering but also it has outlier rejection capabilities and can give insights towards the degree under which the participation of a row or a column is most effective. Several previous possibilistic approaches are based on computing the zeros of an objective function. However, they are sensitive to their input parameters and initial conditions beside that they don't allow constraints on biclusters. This paper proposes an iterative algorithm that is able to produce k-possibly overlapping semi-possibilistic (soft) biclusters satisfying input constraints. The proposed algorithms basically alternate between a depth-first search and a breadth-first search to effectively minimize the underlying objective function. It allows constraints, applicable to any acceptable (dis)similarity measure for the type of the input dataset and it is not sensitive to initial conditions. Experimental results show the ability of our algorithm to offer substantial improvements over several previously proposed biclustering algorithms. © Springer-Verlag Berlin Heidelberg 2012.