Outlier detection and clustering of fifth-generation wireless channel model datasets

Abstract

The fifth-generation (5G) wireless communications system offers faster data rates, lower latency, and more interconnecting devices. Various 5G channel models were developed to study its stochastic characteristics before implementation. These channel models generate multipath components that are grouped into clusters. The multipath clusters serve as datasets in multipath clustering. The clustering results are then used to examine the propagation properties of the 5G system. However, datasets are prone to outliers. They tend to affect clustering accuracy. Hence, this study clusters the datasets generated by the channel models using five clustering approaches, removes the outliers using mean-shift outlier detection, and clusters the datasets free of outliers again using the same clustering algorithms. Outlier detection shows that 5G channel model datasets contain noise, and outlier removal improves the modeling characteristics, as demonstrated by enhanced clustering accuracy. Results show that most of the outliers are detected in the 2×SD threshold. The removal of the outliers using the said threshold increased the clustering accuracy of K-means and AC-Single in Semi-Urban B1 LOS multiple links by 78.85% and 55%, respectively, and DBSCAN in Semi-Urban B2 LOS multiple links by 57.14%. Outlier detection and removal also work well with 5G channel model datasets.