Abstract:

A novel approach using density pointer is proposed to initialize cluster centers. The density pointer (DP) algorithm takes the geometric centers of grid cells as symmetrical centers. With the interconnections between these centers and the vertices of grid cells, DP partitions traditional rectangular-like grid cells into hyper triangle-like subspaces. The density differences between hyper-triangle subspaces and the corresponding subspaces of their neighborhoods are considered to define density pointers. After that, DP detects density pointers to calculate the aggregation factors of dense cells and then takes the gravity centers of the cells with larger local aggregation factors as initial cluster centers. Experiments on both public and real datasets show that DP is helpful to find cluster centers near to real centers quickly and effectively. Moreover, the running time of DP is linear with respect to the number of instances, the number of grid cells and the dimensions.

Key words: density pointer, aggregation factor, cluster centers, initializing