Obtaining the Initial Centroids Based on the Most Dense Colonies in the k-Means Algorithm

In this paper, the standard k-means algorithm has been improved in terms of the initial cluster centers. In the standard k-means method, the first centroids are chosen randomly from members of a dataset, but in the proposed plan, the nearest data points are extracted based on the Euclidean distance and a group is created. The average of all points in a class creates a further element which is distinguished as a candidate for an early centroid. In the next step, the data points which are selected are disconnected from the dataset temporally and another attempt is made to obtain the next compact body and, therefore, another possible candidate and so on. The number of candidates should be four times larger than the number of given classes in order to obtain superior performance and to undoubtedly determine the candidates belonging to all parts of the dataset distribution. Next, the initial centroids are selected from the available candidates based on the number of classes and the geometric assessment with respect to the furthest candidates. The initial centroids are applied to the k-means algorithm and the final clustering is calculated with a few artificial data points and also some common datasets for three classes, and finally the findings are compared with those of the standard k-means technique. The experimental results show that the epochs are lower, the algorithm is robust to trapped in local minima, and also the accuracy is improved properly