Shrinkage k-Means: A Clustering Algorithm Based on the James-Stein Estimator

Abstract

In this work, we propose Shrinkage k-means (Sk-means), a novel variant of k-means based on the James-Stein estimator for the mean of a multivariate normal given a single sample point. We evaluate Sk-means on both synthetic and real-world data. The proposed method outperformed standard clustering methods and also the existing method based on k-means which uses the James-Stein estimator. Results also suggest that Sk-means is robust to outliers.