Abstract: Aiming at the initial clustering center of k-medoids may be too near, under-represented, or poor stability, an improved k-medoids algorithm is proposed. The ratio of sample sets average distance and samples average distance is treated as the density of sample parameters, the number of candidate representative points in the high density point set is simplified, the product of maximum distance method is adopted to choose K samples with high density and long distance as the initial clustering center, both of the representative and dispersion of the clustering center are considered also. Experimental results on the UCI data set show that compared with the traditional K-medoids algorithm and the other two improved clustering algorithms, the new algorithm not only has more accurate clustering results, but also has faster convergence speed and higher stability.

Key words: density, initial cluster center, K-medoids, absolute error