Understanding Societal Phenomena and Network Operations with Mobile Metadata

Abstract

Mobile networks play a fundamental role in modern society, enabling connectivity for billions of users worldwide, supporting digital services, facilitating alerts and coordination during extreme events, and continuously evolving to accommodate new use cases. While carrying out their primary function of supporting telecommunication services, mobile networks generate vast amounts of metadata, which serve as a valuable source of insights into human mobility, urban dynamics, and social phenomena, and for improving the operation of the telecommunication system itself.

In this thesis, we explore various applications of mobile network metadata to social analyses and network infrastructure sustainability while addressing key problems associated with using such metadata, like geolocation and availability.

We start this thesis by developing new methodologies and models to evaluate and improve the spatial accuracy of mobile metadata analysis. We show how traditional approaches to spatial mapping of mobile metadata, such as Voronoi tessellation, suffer from significant limitations, often misrepresenting the proper geographical distribution of network metadata like traffic demands. Our quantitative assessment, based on a unique ground truth dataset provided by Orange, reveals that Voronoi cells achieve only modest accuracy in representing the geographical distribution of network metadata, with a median F-score of 0.56 at best. To address these limitations, we introduce two novel methodologies. VoronoiBoost enhances the traditional Voronoi-based approach through an adaptive scaling mechanism that incorporates non-uniform probability distributions and allows for overlapping coverage areas. This Gradient Boosting Regression model improves mapping quality by an average of 44% over standard Voronoi cells. Taking a more advanced approach, DeepMEND employs a teacher-student framework where a complex probabilistic inference model generates soft labels to train a deep neural network, further reducing median errors by 56.8% compared to Voronoi tessellation and 33% over VoronoiBoost.

Beyond geolocation improvements, we address the challenge of democratizing access to mobile network metadata through the NetMob23 dataset, released via the NetMob 2023 Data Challenge. This dataset provides high-resolution traffic demand data for 68 individual mobile services across 20 major metropolitan areas in France over a continuous 77-day period. A distinguishing feature is its exceptional spatial resolution of 100 m² grid cells, spanning over 870,000 cells with time series of traffic demand at 15-minute intervals. The Data Challenge engaged participants from more than 100 institutions worldwide, fostering new research directions and resulting in over 25 publications across various venues.

We explore novel applications across several domains based on these methodological and data contributions. First, we investigate the relationship between mobile service demand and electoral outcomes in the 2019 and 2024 French European Parliamentary elections. Our Dirichlet regression models demonstrate that mobile service consumption patterns are strongly associated with voting preferences, sometimes even more than traditional socioeconomic variables. Including mobile service demand significantly improves model explanatory power, with average variance explained increasing by up to 20% and 70% in 2019 and 2024, respectively. Specific findings reveal that Facebook usage consistently correlates with higher vote shares for Rassemblement National. At the same time, Instagram demand is negatively associated with far-right and far-left parties but positively linked to centrist and progressive parties.

We further explore mobile traffic demand for understanding urban dynamics through tensor decomposition techniques. Our analysis reveals that the Tucker decomposition provides a richer set of factors for identifying complex relationships between spatial regions, temporal patterns, and service usage. This methodology distinguishes between residential, commercial, and commuting zones based on service consumption patterns, offering insights into urban functional areas.

Complementing this work, we examine the impact of COVID-19 response measures on mobile service usage in France over seven months spanning multiple lockdowns and curfews. Our findings show that strict lockdowns significantly reduced overall mobile traffic, while milder restrictions like curfews allowed for gradual recovery. Crucially, examining hundreds of services independently reveals diverse behavioral patterns: home-centric services saw increased usage during lockdowns, while mobility-related apps experienced significant declines.

Finally, we address sustainability in mobile networks by evaluating five energy-saving policies implemented in a production network. Our analysis of cell sleep strategies in regions with different population densities reveals significant variation in energy savings, with the most aggressive approach achieving reductions of 34.5% and 30.2% in dense and sparse areas, respectively. However, considering the entire network, total energy savings are below 10%, indicating substantial room for improvement. We also identify essential trade-offs between energy efficiency and service quality, providing insights for more effective energy-saving approaches.

In conclusion, this thesis makes significant contributions through methodological innovations, data democratization initiatives, and novel applications across political analysis, urban studies, and network sustainability. Looking ahead, short-term perspectives include refining geolocation methods, expanding data accessibility, deepening political and social analyses, evaluating 5G adoption, and advancing sustainable network management strategies. Longer-term challenges involve adapting to new user interfaces and devices, addressing competition from satellite constellations, and developing explainable AI-driven autonomous networks.

These contributions collectively advance our understanding of mobile networks not just as communication infrastructures but as large-scale sensor systems that provide valuable insights into human behavior, inform policy decisions, and support more sustainable technological development. Thus, they ultimately contribute to more resilient, inclusive, and sustainable digital ecosystems.