On mobile usage data analysis, data-driven network optimization and data synthesis
View/ Open
Date
30/11/2021Author
Singh, Rajkarn
Metadata
Abstract
Recent advancements in mobile systems have led to a rapid proliferation of smart phones
and smart devices into our daily lives. This has given rise to a plethora of new services
and applications for commercial businesses as well as personal end user consumption.
Moreover, the rate at which mobile Internet data is consumed has also increased many-fold in the recent years. Therefore, in order to optimally serve mobile data consumers as
well as improve the underlying system performance, understanding and characterizing
mobile traffic data becomes essential.
In the recent past, mobile networking domain has witnessed several technological
innovations including cloudification and virtualization of radio access networks (RAN),
advanced resource orchestration via network slicing, massive MIMO antennas, millimeter wave etc. These technological advancements, however, present a paradigm shift
from the traditional cellular networks and bring a higher complexity in the management
and orchestration of mobile networks due to an increased number of decision variables.
Therefore, it becomes important to develop efficient, scalable, traffic-aware solutions
for the optimization of mobile network performance.
In this thesis, we take a step forward towards mobile network data analysis and
network performance optimization. We start by experimenting with a nationwide traffic
dataset and build our understanding of mobile traffic usage characteristics at the services
(i.e., mobile apps) level. We derive the correlation between service usage and the
underlying spatial demographic features.
Further, using a data-driven approach, we solve the problem of network resource
orchestration in the virtualized RAN (vRAN) setting. We propose a distributed and scalable heuristic solution, GreenRAN, to minimize the overall vRAN energy consumption.
A major factor limiting this traffic-driven analytics research is the scarce availability
of real-world networking datasets. Access to these datasets is generally restricted either
due to the challenges involved in large-scale database access management or due to
the privacy issues. To lower this dataset access barrier, we develop a tool, MTGAN, for
generating synthetic high-fidelity mobile traffic datasets. Our synthetic traffic generator
is a deep neural network based on Conditional Generative Adversarial Networks.
Finally, we also study and compare different state-of-the-art mechanisms for privacy
preserving data publication. We develop a metric, STRAP, to evaluate user privacy
provided by different privacy preserving mechanisms and compare them under a common
measure of privacy as determined by STRAP.