Abstract: Aiming at the problem that the number of experiment had an inaccurate effect on the square of errors when the K-
means++ algorithm was used to select the initial clustering center to calculate the sum squared error, we proposed a PK-means++ algorithm. The results show that the sum squared error after clustering is more stable than the original K-means++ algorithm under the same K-value when the algorithm clusters the scattered data, so the stability of random experiment value is better guaranteed.

Key words: clustering analysis, K-means++ algorithm, probability, sum squared error