ABSTRACT
Stability of a biclustering solution for non-deterministic algorithms is a central issue which highly influences the ability to interpret the results of a biclustering analysis. Filippone et al. (2009) argued that the stability of biclustering algorithms can be affected by initialization, parameter settings, and perturbations such as different realization of random noise in the dataset. The initialization affects mostly the algorithms relying on local search procedures. In such cases, the outcome of a biclustering procedure is highly dependent on the choice of initial values. Most of the algorithms try to overcome this drawback by using a large number of initial values (seeds) and then finding the best biclusters in the final output set (Murali and Kasif, 2003; Bergmann et al., 2003; Shi et al., 2010). In practice, running a chosen biclustering algorithm several times on the same dataset will give a different output. It remains still up to the analyst to choose the most reliable biclusters or the most robust biclusters given the specification of initial values and the parameters setting.
