Grid-DPC: Improved density peaks clustering based on spatial grid walk

Bo Liang, Jianghui Cai, Haifeng Yang

Abstract: Traditional clustering methods need to find the initial centers first. A reasonable cluster center can improve the efficiency and accuracy of the algorithm. However, finding centers is not an easy task. It often needs much calculation and easily falls into local optimal points. In allusion to the problem, an improved density peaks clustering algorithm based on spatial grid walk (Grid-DPC) is proposed. Grid-DPC uses a spatial grid walk method to determine the initial cluster centers, avoiding the traditional method requiring multiple iterations to optimize the centers and preventing falling into the local optimal. In terms of density definition, Grid-DPC uses grid density instead of the traditional method, which reduces the time cost of finding neighbors. In terms of the random walk strategy, the adjacent higher grid weight direction and Lévy Flight based step length method are adopted to improve the convergence speed of the algorithm effectively. Aiming at the phenomenon of “target loss” that may occur in random walks, the variable neighborhood search method is used to help the algorithm find high-weight grids and prolong the number of steps per walk as long as possible. Simulation experiments show that the algorithm can effectively find clustering centers, complete the task of clustering arbitrarily distributed data, and has high efficiency for massive data.