Harris hawks optimization algorithm based on elite fractional mutation for data clustering

摘要

The density peak clustering (DPC) algorithm is an efficient clustering algorithm that can automatically find the class center and realize arbitrary shape data clustering. The design of the local density is the core of the DPC algorithm, and the value of the cutoff distance parameter involved in the calculation of the local density has a great impact on the performance of the algorithm. In this paper, based on analyzing the defects of the local density design of the DPC algorithm, we use the cosine similarity and exponential decay function to establish a new method of segmented local density calculation, and build an optimization model for the selection of cutoff distance parameters. A new density peak clustering algorithm (NDPC) is proposed. When solving the model, an improved Harris hawks optimization algorithm (FHHO)based on elite fractional derivative mutation is proposed. Simultaneously, the FHHO-NDPC algorithm combining FHHO and NDPC algorithm is put forward. The FHHO algorithm uses Grünwald-Letnikov (G-L) fractional derivative to correct the elite population which changes with the number of iterations and uses a more random exploration strategy to enhance the exploration performance of HHO algorithm. Therefore, the proposed FHHO algorithm inherits the merits of fractional derivative memory, mends the ability of exploration and exploitation by random exploration strategy, and refrains from the algorithm sinking into local optimum. Two groups of experiments are devised simultaneously to verify the significance and usefulness of the FHHO algorithm and FHHO-NDPC algorithm. Experimental results on the CEC2017 test set show that FHHO has obvious dominant positions in solving high dimensional problems in terms of convergence speed and solution precision compared with other representative intelligent algorithms. The clustering results of twelve representative data sets show that FHHO-NDPC has an excellent clustering performance, which provides a useful reference for the design of large-scale data clustering algorithms.