Abstract: The main idea of NJW is to project data points into feature space and then to cluster using K-means algorithm,while the clustering results is considered as the clustering of the original data points.However,the number of clusters C and scaling parameter σ of Gaussian kernel function greatly influence the clustering performance of NJW algorithm,whilst the fact that K-means algorithm is sensitive to initial cluster centers also influences the clustering result of NJW algorithm.To this end,an improved NJW algorithmis presented referring to DP-NJW algorithm which heuristically determines cluster-center points and the number of clusters according to density distribution of data points,and then implements NJW algorithm to cluster.Note that,the obtained cluster-center points and the number of clusters in the first stage are the initialization of K-means algorithm in the second stage of the proposed DP-NJW algorithm.In the next section,DP-NJW is compared with state-of-the-art clustering algorithms on the given seven public datasets,where DP-NJW provided higher clustering accuracy than NJW on five datasets and even on the other two datasets.DP-NJW algorithm provided better clustering performance than DPC on the given five datasets.In addition,DP-NJW consumed less computing time than the other two algorithms,and this is especially obvious on the larger aggregation dataset.Overall,DP-NJW algorithm is superior to the state-of-the-art clustering algorithms.

Key words: Clustering accuracy, Gaussian kernel function, NJW clustering, Scaling parameter, Spectral clustering