Coverage Prediction for Inter-Frequency Handover using Machine Learning with Aggregated Training Data

Abstract: An important application of Machine Learning (ML) in mobile networks is to predict if a user device has coverage on a frequency other than the current serving frequency. This is a use-case called Secondary Carrier Prediction (SCP). In this thesis we deduce if data across different cells and frequencies can be successfully combined when learning this task, thus reducing the number of models that require training. Aggregation of data involves several challenges, such as different prevalences and varying amounts of available data, but more importantly the possibility of achieving synergies in training by exploiting recurring patterns in data. By using an experimental setup in which models are trained and validated on aggregated datasets it is shown that synergies in fact can be achieved through aggregation. The scalability of this task is improved so that the number of models can be reduced with a factor as large as the number of cells times the number of frequencies, while maintaining similar or improved prediction performance.