Comparison of Clustering Techniques Using an Indoor Measurement At 300 GHz

In recent years, cluster-based channel models comprising clusters or groups of multipath components (MPCs) have found a lot of acceptance in the research community. However, such models are heavily clustering technique dependent where there is a primary lack of uniformity regarding distance measure or the number of initial parameters. In this work, four clustering techniques, using the same power-weighted MPC distance as the distance measure and having the same number of user-specified initial parameters are introduced. The newly evolved techniques with uniform assumptions are next compared in terms of the average Silhouette coefficient (ASC) to assess their clustering quality in terms of cluster compactness using data from an indoor measurement campaign. Since clustering techniques normally do not consider the propagation environment, the relevance of the clusters generated using the introduced techniques is also assessed with the measurement environment using ambiance-embedded spectra and ray tracing simulation. It is observed that the variation of KPowerMeans whose a priori parameters are environment and spectra-dependent performs better both numerically and qualitatively compared with the other techniques. Furthermore, it is observed that independent of the technique there are always clusters that cannot be verified from the environment irrespective of the ASC value. Thus, it is concluded that numerical measures are only indicative and should be used in conjunction with environment-based assessment for conclusive results.