Silhouette coefficient-based weighting k-means algorithm

The classical k-means algorithm utilizes all features of the data equally for clustering. It fails to distinguish the roles played by different features in clustering, resulting in poor adaptability. In this paper, we propose a Silhouette coefficient-based weighting k-means algorithm (WKBSC) that automatically adjusts feature weights when clustering. In this algorithm, the optimal clustering is transformed into an optimization problem, which aims to maximize the Silhouette coefficient of the outputting clusters. The feature weights are updated during the iterations of k-means. Compared with eight clustering methods on ten datasets, our method achieves the highest score. We also insight into four datasets and reveal the factors that influence the performances of our algorithms. In summary, the experiment results verify that the proposed algorithm can automatically adjust feature weights to reflect the importance of features in clustering, and can be used for clustering complex real data.