Network Service Disruption upon Natural Disaster: Inference Using Sensory Measurements and Human Inputs

Natural disasters cause large-scale network service interruption which corresponds to unreachability of networks. This problem relates to how networks respond under extreme conditions, and it is neither well-studied nor well-understood. To infer network service disruption, challenges arise, i.e., how to use heterogeneous data that include sensory measurements and human inputs? This work shows an important role of data mining and machine learning in inferring large-scale network service disruption upon Hurricane Katrina. We present a joint use of large-scale sensory measurements from Internet and a small number of human inputs for effective network inference. Specifically, data mining includes (a) unsupervised learning, i.e., clustering and feature extraction of sensory measurements and (b) semi-supervised learning of both sensory measurements and human inputs. The approaches are evaluated on network service disruption induced by Hurricane Katrina at subnet level. Our result shows that clustering reduces the spatial dimensionality by 81%, and sensory measurements are temporally extracted down to two features. The subnet statuses inferred by the classifier derived from semi-supervised learning show interesting facts of network resilience and provide the spatial and the temporal maps of network service disruption that can be used to assist disaster response and recovery. To our understanding, this is the first work of data mining and machine learning using sensory measurements and human inputs for inference of large-scale network service disruption upon a large-scale natural disaster.