Machine Learning as a Service for Beyond 5G Networks

Abstract

5G Network is designed to support numerous services. One of the major challenges faced by operators is to determine a service traffic pattern and dynamically apply machine learning (ML) models. These learning models are aimed to detect, predict and mitigate the network problems. Another critical problem is to optimally allocate the limited ML resources to exponentially increasing devices and number of base stations. Existing solutions utilize the limited ML resources for training and inference using manual methodologies. Empirical studies show that network traffic pattern is periodical and dynamic and hence the manually configured ML models with limited ML resources may severely degrade the network performance. When deploying ML based use cases for 5G services, the ML models need to be trained for all the cells in the network. This will put severe strain on the ML related processing due to limited ML resources allocated. To overcome aforesaid challenges, this research paper proposes an ML-as-a-Service framework that intelligently provides the ML package based on the service profile, region-wise resource usage pattern, KPI list of optimization set by operator, and current ML resource usage in the network. We dynamically and automatically group the cells across Base Stations (BS) using statistical correlation algorithms and train only a master cell in each group using the best training model, which is automatically picked up by our proposed framework. We apply the trained model of master cell to remaining BS within the group. We test our solution on live operator data (both 4G and 5G across different network KPIs) and show the effectiveness of our framework based on different proposed ML KPIs. Our solution saves on an average around 66.15% of CPU time, 60.13% of CPU memory, 63.62% of CPU Utilization and 182 ML trainings across different KPIs for 4G network KPI data without compromising the ML prediction accuracies.