Clustering Performance Analysis of the K-Medoids Algorithm for Improved Fingerprint-Based Localization

Abstract

Fingerprint-based localization, which uses received signal strength (RSS) measurements from spatially deployed wireless access points (APs), is a popular technique for indoor positioning. The size of the fingerprint database has a significant impact on the accuracy of localization. The higher the density of the fingerprint database, the more accurate the localization, but the longer the localization time. Clustering is one of the techniques used such systems to improve localization accuracy and reduce localization time. To cluster fingerprints, the majority of clustering techniques employ a distance-based fingerprint similarity metric. However, the choice of distance metric has a significant impact on the performance of the clustering algorithm. Using four publicly available RSS-based fingerprint databases, this paper investigates the clustering performance of the k-medoids algorithm using six distance metrics, namely Euclidean, Manhattan, cosine, Mahalanobis, Chebyshev, and Canberra distance. Using the silhouette score as a performance metric, the cosine and Euclidean distance metrics outperform the others, with the highest silhouette score values of about 0.38, 0.43, 0.34, and 0.31 on the SEUG_IndoorLoc, IIRC_IndoorLoc, MSI_IndoorLoc, and IPIN_2019_PIEP_UM databases, respectively. It demonstrates that on these four databases, using Euclidean distance as well as the angle between fingerprint measurement vectors is the best option for generating efficient clusters that will result in high localization accuracy and low localization time.