Title:
Data Driven Approaches for Traffic Prediction and Topology Inference in Communication Networks
Abstract:
Modern communication networks require accurate, data-driven methods to understand both how traffic evolves over time and how network structure affects end-to-end performance. This dissertation develops learning-based frameworks for two complementary problems in network management: traffic matrix prediction and network topology inference.
The first part of the dissertation focuses on traffic matrices, which describe the traffic exchanged between source-destination pairs in a network. Because Internet traffic is often bursty, nonstationary, and heterogeneous across flows, models trained directly on raw data can miss the underlying temporal structure. To address this, this work develops a preprocessing framework based on multivariate singular spectrum analysis and studies time-series clustering as a scalable prediction strategy. Together, these methods improve traffic matrix prediction by extracting shared temporal structure and decomposing the forecasting task into smaller, more manageable subproblems.
The second part of the dissertation addresses topology inference from end-to-end delay measurements collected only at boundary nodes. Since multiple physical topologies can produce similar observations, exact recovery is inherently ambiguous. This work reformulates topology inference as a supervised classification problem over topology equivalence classes and shows that accurate inference is possible for moderate network sizes.
Overall, this dissertation demonstrates how structure-aware learning can improve both temporal prediction and structural inference in communication network.
Research Advisor:
Prof. Alex Wyglinski
ECE Department, WPI
Committee Members:
Prof. Bo Tang
ECE Department, WPI
Prof. Charlotte Fowler
Mathematical Sciences, WPI
Prof. Randy Paffenroth
Mathematical Sciences, WPI