Initialization of cluster refinement algorithms: a review and comparative study

Maxmin Data Range Heuristic-Based Initial Centroid Method of Partitional Clustering for Big Data Mining

“…Various research contributions identified greedy methods for the optimization of the initial centroid. Some greedy based initial centroid approaches (Fränti & Sieranoja, 2019) are the greedy technique (He et al, 2004), subsampling (Celebi et al, 2013), and repeated strategy (Bradley & Fayyad, 1998). He et al (He et al, 2004) used the greedy method for optimization of the cluster distance for achieving the cluster objectives through the KM clustering.…”

“…He et al (He et al, 2004) reviewed and measured the random centroids, distance optimization, and density estimation for the initial centroid methods basis of the quantitative property. The Forgy (Forgy, 1965) and MacQueen (MacQueen, 1967) methods are categories under the random centroids, the Simple Cluster Seeking (SCS) (He et al, 2004) method and other greedy variant methods are categories under the distance optimization, and the Kaufman method (Gentle et al, 1990) and other Maximin variant methods are categories under the density estimation. This contribution observed that the initial centroid methods abandoned the cluster separation and considered cluster compaction during the KM algorithm optimization.…”

“…Based on the existing research perspective, and comparative analysis of (Celebi et al, 2013;Fränti & Sieranoja, 2019;He et al, 2004;Peña et al, 1999;Steinley & Brusco, 2007) finds which initial centroid methods are suitable for the KM algorithm. Here, this paper examines which initial centroid methods are achieved better performance under the big data environment through high volume data processing capability, convergence speed, time complexity, scalability, computation efficiency, memory efficiency factor.…”

“…Here, this paper examines which initial centroid methods are achieved better performance under the big data environment through high volume data processing capability, convergence speed, time complexity, scalability, computation efficiency, memory efficiency factor. Pros and cons examinations of the initial centroid methods are shown in table 1 for big data clustering through the discussed literature and comparative analysis (Celebi et al, 2013;Fränti & Sieranoja, 2019;He et al, 2004;Peña et al, 1999;Steinley & Brusco, 2007) using random centroid, random partition, repeated heuristics, maxmin/distance optimization, greedy heuristics, sort heuristics, projection heuristics, density heuristics, and split heuristics categories. (Fahad et al, 2014;Fränti & Sieranoja, 2019) This parameter identified data processing capability in the massive datasets for achieving the initial centroid.…”

“…Convergence Speed (CS) (He et al, 2004;Peña et al, 1999) This parameter identified the number of iterations for achieving the centroid. The best initial centroid method presents the higher convergence speed rate and high convergence speed that are obtained by fewer iteration.…”