An Improved Clustering Algorithm Based on Cluster Weight Coefficient

Clustering algorithms are typical unsupervised machine learning algorithms, which are widely used in many fields. As a popular clustering algorithm, K-means has a good performance, but it has difficulties in determining initial clustering centers and the number of clusters. Canopy algorithm is often used to support K-means as a coarse clustering process. The results obtained from Canopy are set as the initial clustering centers of K-means. However, the thresholds of Canopy are often randomly selected, which makes the clustering result inaccurate. In this paper, we improve the threshold selection of the traditional Canopy and the criterion function of the traditional K-means. As a result, a novel Cluster-Weight-based Canopy-Kmeans (CWC-Kmeans) is proposed. In order to evaluate the proposed algorithm, comparison experiments are conducted with UCI standard datasets and bearing nonstandard dataset. By typical clustering evaluation indices, CWC-Kmeans has better clustering results than the traditional Canopy and the traditional K-means. Thus, the superiority of CWC-Kmeans is proved.