This thesis seeks to formulate concepts and develop methods that facilitate the mining of urban big mobility data. Specifically, the aim of the formulations and developed methods is to identify and predict certain events that occur as a result of urban mobility. This thesis, studies unexpected gathering and dispersal events.
A Gathering event is the process of an unusually large number of moving objects (e.g. taxi) arriving at the same area within a short period of time. It is important for city management to identify emerging gathering events which might cause public safety or sustainability concerns. Similarly, a dispersal event is the process of an unusually large number of moving objects leaving the same area within a short period of time. Early prediction of dispersal events is important in mitigating congestion and safety risks and making better dispatching decisions for taxi and ride-sharing fleets. This thesis solves the problems of early detection and forecasting of gathering and predicting dispersal events.
Prior work to detect gathering events uses undirected patterns which lack the ability to specify the dynamic flow of the traffic and the destination of the gathering. Forecasting gathering events is a predictive approach as apposed to descriptive approaches of detection. This thesis is the first to use destination prediction to forecast gathering events. Moreover, the presented destination prediction technique relaxes independence assumptions of related work and addresses the resulting challenges to achieve superior performance. Literature of dispersal event prediction solves this problem as a taxi demand prediction problem. Those methods aim at predicting the regular pattern and are unable to predict rare events.
This thesis presents the SmartEdge Algorithm for early detection of gathering events. SmartEdge outputs a gathering footprint that specifies gathering paths and gathering destination. To forecast gathering events, this thesis presents DH-VIGO, which uses a dynamic hybrid model to forecast rare gathering events ahead of the time. Comprehensive evaluations using real-world datasets demonstrate meaningful results and superior performance compared to baseline methods.
To predict dispersal events, this thesis uses a two-stage framework based on survival analysis called DILSA+, to predict the start time of the event and an event demand predictor to predict the volume of the demand in case of a dispersal event. Extensive evaluations on real-world data demonstrate that DILSA+ out-performs baselines and can effectively predict dispersal events.