Gravitational clustering algorithm based on mutual K-nearest neighbors

Abstract

To address the problems of difficulty in determining the truncation distance, single definition of local density and low robustness of non-centroid assignment strategy and chain reaction in density peaking clustering algorithm (DPC), this paper proposes a gravitational clustering algorithm (GMNN) based on mutual K nearest neighbors. The algorithm redefines the similarity metric and local density using the mutual K-nearest neighbor approach. Based on the local gravity model, a two-step clustering strategy is designed to isolate the chain reaction to complete the clustering through the mutual gravity between points and clusters. It is marked by simulation experiments that DG-DPC algorithm is effective for both synthetic dataset and UCI dataset, and the accuracy rate is improved by 31.07%, 45.60%, 50.20%, and 35.5% on average relative to RE-DPC algorithm, DPC algorithm, GAP-DPC algorithm, and DG-DPC algorithm, respectively.