DPC algorithm based on K-reciprocal neighbors and kernel density estimation

Clustering by fast search and find of density peaks (DPC) algorithm is a density-based clustering algorithm that does not require iteration or too many parameter settings. However, it fails to identify cluster centers with low cluster density because the local structure of data is not considered when computing local density. To solve this problem, a DPC algorithm based on K-reciprocal neighbors (KN) and kernel density estimation (KDE), called KKDPC was proposed. The number of KN and local kernel density of data points were obtained using the k-nearest neighbor and KDE methods. The number of KNs and local kernel density were weighted to obtain the new local density. The relative distance of data points was obtained based on their local density, and cluster centers and non-center points were selected based on the decision graph. Experiments were conducted on artificial and real datasets and compared with seven clustering algorithms including DPC, density-based spatial clustering of applications with noise (DBSCAN), K-means, fuzzy C-means clustering (FCM), DPC based on K nearest neighbors (DPC-KNN) algorithm, DPC with nearest neighbor optimization (DPC-NNO) algorithm, and DPC-FWSN algorithm. The performance of the KKDPC algorithm was verified by calculating the adjusted mutual information (AMI), adjusted Rand index (ARI), and normalized mutual information (NMI). The experimental results show that the proposed KKDPC algorithm can accurately identify cluster centers and improve clustering accuracy effectively.