An Improvement of Density Peaks Clustering Algorithm Based on KNN and Gravitation

Clustering by fast search and find of Density Peaks (DPC) is a famous clustering algorithm. The main advantages of DPC are that its clustering process is iteration-free and it is independent from the shape and dimensions of a dataset. However, DPC has two main weaknesses: 1) it is subjective to select the cluster centers from the decision graph since they are selected by a user, and thus some selected centers are not necessarily correct; and 2) if a sample point is put into a wrong cluster, then this error can fast be propagated and thus makes many sample points be wrongly put into the cluster. To overcome the above weaknesses, the paper improves DPC (called IDPC). We propose a new density formula combined with the idea of Gravitation and KNN that can make the local densities of sample points in dense and sparse areas have more obvious separability. IDPC does not require users to manually select cluster centers from a decision graph. Experiments on 15 datasets show that our clustering algorithm overall outperforms the state-of-the-art ones including ADPC-KNN, DPC, DPCSA and DPC-KNN.