Study of network-service disruptions using heterogeneous data and statistical learning

Communications comprises a key infrastructure that supports every aspect of our daily lives. In the past ten years, large-scale disturbances have proven to cause extensive damage to the communications infrastructure and require millions of dollars to repair the damage. Considerable attentions in the prior work have focused on assessing network damages after natural or man-made disasters. However, network-disruption responses, i.e., how the disruptions occur depending on social organizations, weather, and power resources, have been studied little.;The objective of this research is to study network-service disruptions caused by large-scale disturbances with respect to (1) temporal and logical network, and (2) external factors such as weather and power resource, using real and publicly available heterogeneous data that are comprised of network measurements, user inputs, organizations, geographic locations, weather, and power outage reports.;In this study, network-service disruptions are studied at the subnet level. A subnet is a collection of connected computer devices generally owned by an organization. An unreachability of a subnet occurs if the Internet traffic can no longer route to this subnet.;Network-service disruptions caused by Hurricanes Katrina in 2005 and Ike in 2008 are used as the case studies. First, the identification of subnet unreachability is developed by applying unsupervised- and semi-supervised learning to large-scale network measurements and user inputs.;The network-disruption responses are studied with respect to temporal and logical network dependence. It is found that temporal dependence also illustrates the characteristics of logical dependence. Temporally dependent subnets became unreachable within organization, cross organization, and cross autonomous system. The comparison of subnet unreachability between Ike and normal operations illustrates that subnet unreachability due to Hurricane Ike is indeed anomalous.;In addition, subnet unreachability is analyzed with respect to the storm characteristics. The storm path and coverage are reconstructed from the storm data, and the times and probabilities when the storm coverage overlapped subnet regions are computed. As a result, it is found that subnet unreachability and the storm are weakly correlated.;The weak correlation between subnet unreachability and the storm provides the motivation to search for what exactly caused subnets to become unreachable. We contacted organizations who own unreachable subnets to learn about their actual root causes. Finding root causes of disrupted networks from the subnet owners is challenging since such information is proprietary to organizations. Six out of seven organizations reportedly experienced network disruptions due to power outages or the lack of power generators.;Using power outage data obtained from the Public Utility Commissioner of Texas, the dependence of subnet unreachability on power outages is studied. The network data is aggregated to the same scale as the power outage data. The observations and correlation illustrate that subnet unreachability and power outages are strongly correlated.;The information sharing potentially can be used to improve the state of the art towards studying network-service disruptions. We explore more sharing information resources in weather, power, Internet service providers, daily lives, and emergencies. The best-fit data needed for the dependence study of network disruptions caused by large-scale disturbances are also presented.;This contribution of this thesis is the empirical study of network-service disruptions caused by large-scale disturbances using real and publicly available heterogeneous data and statistical learning. We incorporate network-, weather-, and power-related data into the analysis of network-service disruption caused by large-scale disturbances with respect to temporal and logical network, and external factors such as weather and power resources.