Event Analytics via Discriminant Tensor Factorization.

Analyzing the impact of disastrous events has been central to understanding and responding to crises. Traditionally, the assessment of disaster impact has primarily relied on the manual collection and analysis of surveys and questionnaires as well as the review of authority reports. This can be costly and time-consuming, whereas a timely assessment of an event’s impact is critical for crisis management and humanitarian operations. In this work, we formulate the impact discovery as the problem to identify the shared and discriminative subspace via tensor factorization due to the multi-dimensional nature of mobility data. Existing work in mining the shared and discriminative subspaces typically requires the predefined number of either type of them. In the context of event impact discovery, this could be impractical, especially for those unprecedented events. To overcome this, we propose a new framework, called “PairFac,” that jointly factorizes the multi-dimensional data to discover the latent mobility pattern along with its associated discriminative weight. This framework does not require splitting the shared and discriminative subspaces in advance and at the same time automatically captures the persistent and changing patterns from multi-dimensional behavioral data. Our work has important applications in crisis management and urban planning, which provides a timely assessment of impacts of major events in the urban environment.