Human-in-the-Loop Predictive Analytics for Incident Detection in Smart Transportation Systems

Abstract

The main objective of the PhD thesis is the application of data analytics and artificial intelligence for the automatic detection of incidents in intelligent transportation systems. The thesis proposes a predictive analytics framework and comprehensive methodology to support the entire incident detection process, from data collection to real-time detection and expert validation. It combines artificial intelligence algorithms and predictive methods, as well as automated machine learning (AutoML), to predict both planned and unplanned incidents. Explainability tools, such as LIME and SHAP, are used to explain model decisions to human operators. This enhances confidence in prediction systems, which are typically considered 'black boxes', and makes predictions easier for non-experts to comprehend and act upon, in many cases in a proactive manner. Additionally, the concept of human-in-the-loop (HITL) intervention is incorporated into the methodological process, enabling experts to supervise and correct automatic predictions in real time. This improves both the accuracy of the predictions and the cooperation between human stakeholders and automated systems and algorithms. Corresponding experiments demonstrate the crucial contribution of experts to improving the system's performance dynamically over time. Moreover, in the context of the thesis, an extensive literature review study is conducted on predictive data analytics, automatic incident detection systems, AutoML, explainability, and HITL technologies. The methodologies and systems developed thus far are analysed, followed by the development of the proposed method. Subsequently, an information system is developed to enable the application of the proposed method in urban environments, which is then applied to two different cities, Athens and Antwerp, to evaluate and compare its performance in different contexts.