In complex mobility systems, the widespread presence of spatiotemporal anomaly patterns poses substantial challenges to effective governance and decision-making. A notable example of this challenge is evident in traffic anomalous incidents detection, where the combination of low accuracy in anomaly detection and poor scenario generalization performance significantly impacts the overall performance of anomaly detection. This paper introduces a prototype augmentation-based framework tailored for spatiotemporal anomaly detection in the context of smart mobility system. This framework utilizes prototype augmentation technique to enhance the diversity of anomaly patterns, ensuring that the integrity of the original anomaly information is preserved. Essentially, the prototype augmentation-based anomaly detector employed in this framework is a hybrid unsupervised-supervised stacking ensemble. It leverages the strengths of unsupervised component learners to generate pseudo dimensions while integrating a supervised meta-detector for evaluating the component learners’ performance across diverse environmental contexts. Additionally, we materialize this framework and assess its performance in detecting anomalous line-pressing incidents. Empirical results demonstrate our framework’s superior accuracy and transferability in detecting anomalous traffic incidents compared to alternative methods using a real-world dataset.

中文翻译：

---

智能出行系统中基于原型增强的时空异常检测


在复杂的移动系统中，时空异常模式的广泛存在对有效的治理和决策构成了重大挑战。这一挑战的一个显著示例是流量异常事件检测，其中异常检测的准确性低和场景泛化性能不佳的结合会显著影响异常检测的整体性能。本文介绍了一个基于原型增强的框架，该框架专为智能出行系统背景下的时空异常检测量身定制。该框架利用原型增强技术来增强异常模式的多样性，确保保留原始异常信息的完整性。从本质上讲，该框架中采用的原型基于增强的异常检测器是一个混合的无监督-监督堆叠集成。它利用无监督组件学习器的优势来生成伪维度，同时集成有监督元检测器，用于评估组件学习器在不同环境上下文中的性能。此外，我们还将这个框架具体化，并评估其在检测异常压线事件方面的表现。实证结果表明，与使用真实数据集的替代方法相比，我们的框架在检测异常交通事件方面具有卓越的准确性和可转移性。