Density peaks clustering by granular computing with label propagation

Abstract

Density peaks clustering (DPC) algorithm identifies the density peaks as cluster centers and forms clusters by assigning each object to its nearest higher local density neighbor. However, DPC suffers from two main issues: its local density estimation is unreliable in complex real datasets, and its cluster assignment strategy has a risk of 'chain error'. To overcome these two issues, this paper proposes a novel density peaks clustering by granular computing with label propagation. At first, each object is coarsened into the fuzzy semantic cell. The local density estimation in DPC can be interpreted from learning fuzzy semantic cells and converted into an optimization problem. We provide an alternative reliable local density estimation method. Subsequently, we allocate an initial label matrix by learned fuzzy semantic cells and then utilize the label propagation algorithm to let the label of each prototype be updated from its k nearest neighbors. The adopted label propagation method can significantly improve the cluster assignment strategy compared to DPC. To validate the proposed method, we conduct comprehensive experiments on various datasets. The results prove that our method outperforms the other state-of-art DPC-related methods.