SOM²CE: Double self-organizing map based cluster ensemble framework and its application in cancer gene expression profiles

Though there exist a lot of cluster ensemble approaches, few of them consider how to degrade the effect of noisy attributes in the dataset. In the paper, we propose a new cluster ensemble framework, named as double self-organizing map based cluster ensemble (SOM2CE) to perform clustering on noisy datasets. SOM2CE incorporates the self-organizing map (SOM) twice into the ensemble framework to discovery the underlying structure of noisy datasets, which applies SOM to perform clustering not only on the sample dimension, but also on the attribute dimension. SOM2CE also adopts the normalized cut algorithm to partition the consensus matrix constructed from multiple clustering solutions, and obtain the final results. Experiments on both synthetic datasets and cancer gene expression profiles illustrate that the proposed approach not only achieves good performance on synthetic datasets and cancer gene expression profiles, but also outperforms most of the existing approaches in the process of clustering gene expression profiles.