Uncovering Latent Patterns In Service-Level Spatiotemporal Mobile Traffic

Personal mobile communication technologies are amongst the most successful innovations of the 21st century. The widespread adoption of mobile services has resulted in an exponential surge in mobile traffic, which effectively mirrors human behavior. Mobile devices maintain a continuous interaction with network infrastructure, and the associated geo-referenced events can be easily logged by service providers, for different purposes, including billing and resource management. This leads to the implicit potential of monitoring a significant portion of the entire population at a minimal cost which no other technology provides an equivalent coverage. In this context, analyzing mobile traffic along space, time, and app dimensions can provide actionable insights for improving user experience, optimizing resource allocation, enhancing security, and driving business success in the mobile app and service industry. In this thesis, we employ a tensor decomposition technique, specifically Tucker decomposition, to reveal latent factors (i.e., patterns) spanning the dimensions of space, time, and applications. We apply the technique to real-world mobile traffic data generated by a variety of mobile applications within Paris, France, during 77 consecutive days for every 15 minutes interval. We obtained 4 temporal factors capturing day-night mobile traffic behavior, working hours, commuting, and weekend patterns. In the dimensions of space and mobile applications, we identified 7 factors each which includes a space factor distinctly differentiating between the city center and the rest of Paris. An exploration of the obtained latent patterns in spatial, temporal, and mobile application dimensions reveals interesting interrelationships in user behaviors.