A Variant of K-Means Clustering through Heuristic Initial Seed Selection for Improved Clustering of Data

Unsupervised clustering algorithms have been used in many applications to group the data based on relevant similarity metrics. K-Means clustering is one of the most widely used clustering techniques owing to its simplicity. Many improvements and extensions have been proposed for this algorithm in view to improve its performance. Out of the various dimensions that have been explored in this regard such as mean computation, centroid representation, initial seed/cluster centre selection and similarity calculation methods, the choice of initial cluster centre is found to have a profound impact in the performance of the algorithm. Existing methods chose the cluster centres either randomly or based on heuristics such as maximum distance property, maximum probability of the squared distance, points with maximum points lying close to it etc. In this paper, a strategy to select relevant initial cluster centres for two-cluster grouping problems is proposed based on the measures indicating the statistical distribution of the data in view to improve the clustering performance in terms of accuracy. These measures include minimum, maximum, median, mean and skew of the data. The algorithm is validated on datasets from UCI repository viz. Balance, BloodDonate, Diabetes, Ionosphere, Parkinsons and Sonar and synthetic datasets. The performance of the proposed algorithm is compared with K-Means and its variants and found to achieve better performance in terms of accuracy. An increase is accuracy of approximately 0.25%-18% is observed across the datasets.