Measures of transport mode segmentation of trajectories

Rooted in the philosophy of point-and segment-based approaches for transportation mode segmentation of trajectories, the measures that researchers have adopted to evaluate the quality of the results (1) are incomparable across approaches, hence slowing the progress in the field and (2) do not provide insight about the quality of the continuous transportation mode segmentation. To address these problems, this paper proposes new error measures that can be applied to measure how well a continuous transportation mode segmentation model performs. The error measures introduced are based on aligning multiple inferred continuous intervals to ground truth intervals, and measure the cardinality of the alignment and the spatial and temporal discrepancy between the corresponding aligned segments. The utility of this new way of computing errors is shown by evaluating the segmentation of three generic transportation mode segmentation approaches (implicit, explicit-holistic, and explicit-consensus-based transport mode segmentation), which can be implemented in a thick client architecture. Empirical evaluations on a large real-word data set reveal the superiority of explicit-consensus-based transport mode segmentation, which can be attributed to the explicit modeling of segments and transitions, which allows for a meaningful decomposition of the complex learning task.