Clustering is the process of grouping data based on their same properties. All the elements in each cluster should be similar [1]. The types of clustering include data mining algorithmic clustering, dimension reduction, parallel clustering, and MapReduce-based clustering [2]. Meanwhile, partitioned clustering is a type of data mining algorithmic clustering integrating different algorithms like K-means, K-modes, K-medoids, PAM, CLARA, CLARANS, and FCM [3].

One of the widely used algorithms for clustering implementation is the K-means clustering algorithm [4], whose usage is very common due to the best performance for big datasets [4,5]. In the standard K-means algorithm, K points are firstly selected as initial centroids, where each centroid represents a cluster. Then all objects of the dataset are assigned to the centroids with the minimum distances. After the allocation of all data items, centroids are recalculated until no further objects change their cluster [6]. Generally, Euclidean distance is utilized for this purpose in most cases. However, the allocation may take maximum because it needs to recalculate the distance mathematical equation during each iteration. Therefore, many mathematical metrics are proposed to improve the distance calculation [7].

Mathematical distance measure metrics play an essential role in improving the result of the K-means algorithm. Thus, three distance metrics, i.e., Euclidean distance, Manhattan distance, and Minkowski distance, are implemented in this study. Besides, the execution time with different cluster numbers is evaluated on different datasets, where 100000, 200000, 300000, 400000 and 500000 2D points are randomly selected as datasets.

Many researchers have improved the efficiency and the performance of the K-means clustering algorithm, including the accuracy, the quality of clusters, and the running time of the K-means algorithm [8,9,10]. Kaur et al. [11] presented an improved variant of standard K-means, which provided the image compression with less running time and more efficiency. Dalal et al. [12] introduced an enhanced version of the K-means algorithm to better select starting points so that to meet an improved local minimum. Over the complete dataset, the number of repetitions also decreased. Two things can influence the idea that was dependent upon the best choice of initial centroids. The first one is the novel iterative method, and the second one is optimization formulation. This technique may be implemented on a lot of clustering problems. The technique was also capable of working with many other data mining techniques to obtain the best clustering results. To evaluate the improved algorithm, different experiments were performed on different datasets. As compared to the standard K-means algorithm, the iterations of the proposed K-mean clustering algorithm were fewer to the best performance.

To overcome the drawbacks of standard the K-means clustering algorithm, Gupta et al. [13] presented an improved algorithm without specifying the number of centroids.

There are different clustering types of data mining algorithms like density-based clustering algorithms, hierarchical-based clustering algorithms, partitioning-based clustering algorithms, grid-based clustering algorithms, and model-based clustering algorithms [14]. In partition-based clustering, one of the famous algorithms is the K-means algorithm [14], which first generates a K number of partitions representing the number of groups and then conducts the iterative allocation process of data elements to the group [6].

Bora et al. proposed an experimental study in Matlab to cluster the iris and wine datasets with different distance measures and observed the variation of the performances [15]. Loohach et al. implemented the K-means clustering algorithm with Euclidean distance as well as Manhattan distance metrics and compared the result in terms of the number of iterations. Their results showed that the number of iterations could be affected by the implementation of different distance metrics [16]. Sajana et al. [3] focused on a keen study of different clustering algorithms, highlighting the characteristics of big data techniques and an overview of various types of clustering. Rathore et al. [17] introduced a new technique to implement a K-means clustering algorithm instead of traditional K-means. First, the quality of clusters was improved by removing outlier elements in a dataset; second, the dataset was split into clusters by using a bi-part method. The results were compared with the traditional K-means algorithm and showed better accuracy by removing the de-efficiency.

To find a point-to-point distance between elements and centroid, different distance metrics that play an important role in K-means clustering are measured to assign these elements to related clusters (i.e., centroids). Three distance metrics are implemented and discussed as follows.

We compared execution time with different numbers of clusters on different datasets. Datasets are randomly selected such as 100000, 200000, 300000, 400000, and 500000 in 2D points. Different distance metrics are implemented to measure the distance between data objects. In this paper, the K-means algorithm is employed by using different distance metrics, whose mechanisms are summarized as follows.

In the experiments, Spyder 3.2.8 was implemented, which is a scientific python development environment and powerful python IDE. All experiments were conducted on a machine consisting of an Intel (R) Core (TM) i5-5300 CPU @ 2.30 GHz with 8 GB RAM. The results were evaluated on different numbers of clusters, such as 4, 6, 8, 10, 12, 14, and 16, using five different datasets. To achieve perfect results, many runs were carried out for each use case, and the running times were measured in milliseconds.

In Fig. 1, the running time is compared with three different mathematical methods on four clusters. Different datasets were split into four clusters. Running time is shown along the y-axis in milliseconds, and the compared different datasets of three different distance metrics are shown along the x-axis.

In Fig. 2, compared the running time of the K-means algorithm with three different distance metrics like Euclidean, Manhattan, and Minkowski. It is observed that Euclidean and Minkowski’s methods take the same time at 500,000 dataset points. But in fewer points, the Euclidean distance metric performs better results as compared to other metrics.

In Figs. 3–7, the running time is compared with three different distance metrics, where different sizes of datasets were splits into 8, 10, 12, 14, and 16 clusters, respectively. Running time is shown along the y-axis in milliseconds, and the compared different datasets of three different distance metrics are shown along the x-axis. It can be seen from experiments that Manhattan Distance performs better for 4, 8, 12, and 14 clusters. Euclidean Distance shows better for 6 and 16 clusters, while Minkowski Distance performs better only for 10 clusters.

One of the most popular clustering algorithms is K-means, where different distance metrics are used to find similar data objects. Distance is measured between every point of the dataset and centroids to assign the nearest cluster. In the experiments, the performances of three different metrics (Minkowski Distance, Manhattan Distance, and Euclidean Distance) were measured and compared in terms of execution time with different datasets and different numbers of clusters, i.e., 4, 6, 8, 10, 12, 14, and 16 clusters. It can be seen from experiments that Manhattan Distance performs better for 4, 8, 12, and 14 clusters. Euclidean Distance shows better for 6 and 16 clusters, while Minkowski Distance performs better only for 10 clusters. Overall Manhattan Distance performs better result. In future work, we will try to extend our approach to another partitioned-based clustering algorithm like K-Medoids, CLARA, and CLARANS.