Scalable and Explainable Deep Neural Network Frameworks for Mobile Traffic Forecasting

Abstract

This thesis investigates how deep neural networks can support reliable mobile traffic forecasting for 5G and future mobile networks. Accurate forecasts allow network operators to anticipate demand, allocate radio and computing resources efficiently, improve service quality, and reduce unnecessary energy consumption. However, despite their predictive power, deep learning models are often difficult to interpret and costly to operate at city scale, where traffic patterns vary across hundreds or thousands of base stations. The proposed methods are evaluated using two real-world mobile traffic datasets: the Telecom Italia Milan dataset and measurements from a production LTE network serving a major European metropolitan area.

To address these challenges, the thesis develops an integrated framework centered on explainability, robustness, and scalability. It introduces DeExp, an explainable AI framework that adapts methods such as Layer-wise Relevance Propagation, Grad-CAM, SHAP, and LIME to spatio-temporal traffic forecasting. DeExp converts high-dimensional attribution scores into compact relevance maps, showing which base stations most influence each prediction. These insights are used to analyze model reliability under traffic perturbations, considering prediction error, overprovisioning costs, and service-level agreement violations. Targeted perturbations at influential base stations can increase prediction error by up to 250% in some scenarios.

The thesis further studies realistic mobility-constrained adversaries that can only move locally across neighboring base stations under a limited movement budget. It proposes and evaluates two strategies: a Greedy Mobility-Constrained Attacker and a Smart Explainability-guided Attacker, which combines DeExp relevance scores with look-ahead decision-making. Finally, the thesis proposes a scalable clustering-based training framework that groups base stations with similar temporal behavior using K-means clustering with Dynamic Time Warping and selects the most informative inputs through explainability-guided relevance scores. This reduces telemetry, input-probe, and computational requirements by up to approximately 81% while maintaining competitive, localized forecasting accuracy.