A Cluster Validity Approach based on Nearest-Neighbor Resampling

We introduce an approach for validating clustering results based on partition stability under a nearest-neighbor resampling. The approach is relatively robust, efficient, and avoids conceptual problems of other common validation strategies. Encouraging results compared to those of subsampling-based consensus clustering are presented for simulated data and (tumor) gene expression benchmark data sets. The proposed method is discussed in view of future applications to unsupervised learning from sample data