A comparative study of efficient initialization methods for the k-means clustering algorithm

• Published in: Expert systems with applications Vol. 40; no. 1; pp. 200 - 210
• Main Authors: Celebi, M. Emre, Kingravi, Hassan A., Vela, Patricio A.
• Format: Journal Article
• Language: English
• Published: Amsterdam Elsevier Ltd 01.01.2013; Elsevier
• Subjects: Algorithmics. Computability. Computer arithmetics; Algorithms; Applied sciences; Cluster analysis; Cluster center initialization; Clustering; Clusters; Collection; Computational efficiency; Computer science; control theory; systems; Data processing. List processing. Character string processing; Exact sciences and technology; Expert systems; k-means; Memory organisation. Data processing; Partitional clustering; Software; Statistical tests; Sum of squared error criterion; Theoretical computing; k-means; Sum of squared error criterion; Partitional clustering; Cluster center initialization; Initialization; Gradient; Non parametric test; Statistical analysis; Cluster; Linear time; K means algorithm; Recommendation; Search algorithm; Gradient descent; Experimental result; Efficiency; Classification; Descent method; Database; Linear complexity; Alternative method; Occupation time; Time complexity
• ISSN: 0957-4174; 1873-6793
• DOI: 10.1016/j.eswa.2012.07.021

► K-means is the most widely used partitional clustering algorithm. ► k-means is highly sensitive to the selection of the initial centers. ► We present an overview of k-means initialization methods (IMs). ► We then compare eight commonly used linear time IMs. ► We demonstrate that popular IMs often perform poorly. K-means is undoubtedly the most widely used partitional clustering algorithm. Unfortunately, due to its gradient descent nature, this algorithm is highly sensitive to the initial placement of the cluster centers. Numerous initialization methods have been proposed to address this problem. In this paper, we first present an overview of these methods with an emphasis on their computational efficiency. We then compare eight commonly used linear time complexity initialization methods on a large and diverse collection of data sets using various performance criteria. Finally, we analyze the experimental results using non-parametric statistical tests and provide recommendations for practitioners. We demonstrate that popular initialization methods often perform poorly and that there are in fact strong alternatives to these methods.