Addressing Sparsity in Paratransit Demand and Cancellation Prediction: A Spatiotemporal Kernel Approach

作为一种新兴的公共交通方式，Paratranite为老年人和残疾人提供方便和负担得起的出行服务。鉴于辅助运输服务的运营成本很高，对未来需求和取消航班的准确和可靠估计至关重要，这些估计可以提供给高级优化应用程序，例如车辆路线、乘务调度和车队管理等。然而，与公共交通服务和流行的按需出行不同，对辅助交通服务的需求在空间和时间上基本上是稀疏的。这种稀疏性(计数数据)限制了传统的基于时间序列的模型和新兴的张量学习和深度学习模型的应用，这些模型往往做出高斯假设。另一个挑战是用不同的时空尺度/分辨率做出可靠的估计，从粗略的区域级别到精细的邮政编码级别，从每天的级别到每分钟。为了应对这两个挑战，在这个Engage项目中，我们计划开发一个新的时空建模框架，将核方法与高斯过程泊松回归模型相结合，以准确可靠地预测稀疏需求。为了估计取消率，我们将根据详细的旅行特征(例如，一天中的时间、旅行时间、入住率)、个人特征(例如，年龄、收入)和其他外部特征(例如，天气条件)开发一个通用的分类模型。我们麦吉尔大学的团队(智能交通实验室)将根据GIRO提供的历史(两年)辅助运输预订/取消数据进行工作，新模型将被集成到HASTUS-OnDemand软件中。鉴于时空预测模型无处不在，我们预计研究成果也将惠及GIRO的其他产品，如公共交通需求预测和邮政需求预测。

As an emerging mode of public transport, paratransit provides accessible and affordable mobility services to the elder and the disabled. Given the high operation cost of paratransit services, it is critical to have an accurate and reliable estimation of future demand and cancellations, which can be fed into advanced optimization applications, such as vehicle routing, crew scheduling, and fleet management, to name but a few. However, unlike mass transit services and popular on-demand mobility, the demand for paratransit service is essentially sparse over space and time. This sparsity nature (count data) limits the application of traditional time series-based models and emerging tensor learning and deep learning model, which often makes Gaussian assumptions. The other challenge is to make reliable estimation with different spatiotemporal scales/resolutions, from coarse regional level to fine postal code level, from daily level to minute by minute. To address these two challenges, in this Engage project we plan to develop a new spatiotemporal modeling framework that integrates the kernel approach with Gaussian process Poisson regression models to predict sparse demand accurately and reliably. To estimate the cancellation rate, we will develop a generalized classification model based on the detailed trip features (e.g., time of day, travel time, occupancy), individual feature (e.g., age, income), and other external features (e.g., weather condition).Our group (Smart Transportation Laboratory) at McGill University will work on the historical (two years) paratransit booking/cancellation data provided by GIRO and the new models will be integrated into the HASTUS-OnDemand software. Given that the spatiotemporal prediction models are ubiquitous, we expect the research outcomes to also benefit other products of GIRO, such as public transport demand prediction and postal demand prediction.