A feasible density peaks clustering algorithm with a merging strategy

Density peaks clustering (DPC) algorithm is a novel algorithm that efficiently deals with the complex structure of the data sets by finding the density peaks. It needs neither iterative process nor more parameters. The density–distance is utilized to find the density peaks in the DPC algorithm. But unfortunately, it will divide one cluster into multiple clusters if there are multiple density peaks in one cluster and ineffective when data sets have relatively higher dimensions. To overcome the first problem, we propose a FDPC algorithm based on a novel merging strategy motivated by support vector machine. First, the strategy utilizes the support vectors to calculate the feedback values between every two clusters after clustering based on the DPC. Then, it merges clusters to obtain accurate clustering results in a recursive way according to the feedback values. To address the second limitation, we introduce nonnegative matrix factorization into the FDPC to preprocess high-dimensional data sets before clustering. The experimental results on real-world data sets and artificial data sets demonstrate that our algorithm is robust and flexible and can recognize arbitrary shapes of the clusters effectively regardless of the space dimension and outperforms DPC.