Practical Machine Learning for Predictions in Mobile Networks
Time: Thu 2025-08-21 14.00
Location: Kollegiesalen, Brinellvägen 8, Stockholm
Language: English
Subject area: Computer Science
Doctoral student: Akhila Rao , Programvaruteknik och datorsystem, SCS
Opponent: Associate Professor Rafael Pasquini, Federal University of Uberlândia, Uberlândia, Brazil
Supervisor: Professor Magnus Boman, Programvaruteknik och datorsystem, SCS; Professor Dejan Kostic, Programvaruteknik och datorsystem, SCS
QC 20250523
Abstract
The use of machine learning (ML) in mobile networks has surged to tackle the challenges arising from their growing complexity, dynamic behavior, and application demands.However, more support through prediction models for key performance metrics that cannot be directly observed at the base station are needed to better adapt network and application parameters to the changing network state. Moreover, the growing adoption of ML models in telecom networks brings practical challenges, like the scarcity of labeled data for training, heterogeneity of hardware, and evolving conditions, all of which pose hurdles to their seamless deployment in real-world scenarios.
In this thesis we investigate ML-driven methods to predict user performance in dynamic environments and address the practical challenges encountered over a model’s life cycle. First, we present supervised ML approaches for predicting key metrics such as round-trip time and one-way delay, using data collected from 5G mmWave testbed networks.Further, we propose a mobile edge-assisted framework for improving quality of service (QoS) and reducing backhaul load for video streaming applications, by prefetching video segments, under backhaul bandwidth and cache storage restrictions. This framework is based on ML predictions of video segment bitrates in a dynamically changing mobile network.
Next, we address domain shifts, particularly from the arrival of new user equipment (UE) types, by applying adversarial adaptation strategies that allow models to adapt without requiring labeled samples from the new domain.Further, we apply self-supervised learning approaches to harness unlabeled data from telecom network datasets. This is shown to improve model accuracy in label-scarce scenarios. Finally, to further mitigate data-collection challenges, we introduce an open-source simulation framework for generating large-scale, customizable datasets. Overall, this thesis contributes to the careful design of ML solutions for predicting key performance metrics, and shows how they can, along with robust data generation, model adaptation, and model pretraining techniques, elevate user experience and network automation in modern telecom networks.