An Empirical Seed Initialization Idea for K-Means Algorithm Inspired by CLIQUE Algorithm

The K-means clustering algorithm gives good clustering results when the initial centroids or seeds are chosen close to actual cluster centers. In this paper an empirical centroid initialization technique has been presented for the K-means clustering algorithm. The technique is inspired by a well-known grid-based clustering algorithm called 'CLIQUE.' The technique evaluates the initial centroids by evaluating validity indices to check their vicinity from actual cluster centers. Experiments on datasets show that the technique obtains good initial centroids for the datasets compared to state-of-the-art K-means++ clustering algorithm.