Comparison Study of Distance Methods in Nonlinear Panel Data Clustering with K-Means Method

Abstract

Cluster analysis is used to group objects based on the similarity of characteristics between objects. Cluster analysis will be applied to nonlinear panel data using Indonesian Coronavirus Disease (COVID-19) data with the aim of grouping Provinces based on the number of active positive cases using the K-means method. The first stage will be a simulation to get the best distance method on nonlinear panel data. The distance method used is the Euclidean, Manhattan, Maximum, Frechet, and Dynamic Time Warping (DTW). The simulation results are obtained after running all distance methods with 36 scenarios from four generation data models, the maximum distance method is the best distance method with a total of 20 highest accuracy values compared to other distance methods. The maximum distance method will be applied to real data. The real data results showed that the optimal number of clusters is formed when three clusters are formed with the value of the Calinski Harabatz (CH) criteria of 143,459. Cluster A has 30 members, Cluster B has three members, while Cluster C has one member from DKI Jakarta Province.