Improving consensus clustering with noise-induced ensemble generation: A study of uniform random noise

Specific to data mining or data analysis in general, noise raises the difficulty for many conventional models to deliver a trustworthy result. Several studies have devoted to adjust existing methods to exhibit a noise tolerance characteristic, while some others rely pretty much on the process of data cleansing prior the analysis process. One way or another, the impact of noise is minimized, thus keeping up the goodness of discovered knowledge. In contrary of these, a few researches have recently reported a benefit of injecting small amount of noise into the data under examination. Given such an insight, the paper introduces an initial and unique study of employing noise in the process of cluster ensemble generation. This noise-induced strategy is to deliver data perturbation that can be coupled with general generation methods like homogeneous ensemble of k-means and different number of clusters. In a nutshell, multiple data matrices are created from the original data, each of which possesses salt-and-pepper noise locations and uniform-random noise values. This may yield different cluster structures, hence the diversity within an ensemble. Based on the empirical investigation with nine benchmark datasets, the aforementioned approach has shown potential with improved clustering performance comparing to basic generation methods.