Iterative Screening Density Peak Clustering

Density peak clustering (DPC) is popular in unsupervised clustering due to its intuitive theory and the absence of pre-defined cluster numbers. However, traditional DPC relies on a static single computation strategy. It uses all raw samples to calculate global density and relative distance without iterative correction. Low-density samples in real data distort the density distribution, weaken the peak characteristics of real clustering centers, lead to incorrect center selection and disordered point allocation, and damage the stability and accuracy of clustering. To address this limitation, we propose an Iterative Screening Density Peak Clustering (IS-DPC). A new local density based on binary selection is designed to suppress interference from low-density samples at the computational source. In addition, the iterative mechanism transforms static single-round calculations into dynamic progressive updates. Experiments on real datasets have shown that IS-DPC outperforms several advanced DPC variants in most cases.