Initializing k-means clustering algorithm using statistical information

Abstract

K-means clustering algorithm is one of the best known algorithms used in clustering; nevertheless it has many disadvantages as it may converge to a local optimum, depending on its random initialization of prototypes. We will propose an enhancement to the initialization process of k-means, which depends on using statistical information from the data set to initialize the prototypes. We show that our algorithm gives valid clusters, and that it decreases error and time.