Pattern Recognition of RBS Configuration Topologies in Radio Access Networks

Abstract

This thesis explores how graph-based representations for cluster evaluation may be applied on configuration topologies of radio base stations. It presents experiments performed using aWeisfeiler-Lehman subtree kernel and a graph embedding method based on graph edit distance computations. These techniques are compared to the performance of simpler baseline methods. Experiments are performed on a data set which have been categorised at two levels of detail by RBS configuration experts at Ericsson, and evaluation is performed using both intrinsic and extrinsic metrics. The Weisfeiler-Lehman subtree kernel shows promising results and has a reasonable runtime when executed on the provided data set. The graph embedding technique, on the other hand, gives less promising results even when compared to the baseline methods.