Verbal Explanations of Spatio-Temporal Graph Neural Networks for Traffic Forecasting

Intelligent Transportation Systems (ITS) built using Deep Neural Network (DNN) models offer an effective solution for handling short-term traffic flow, which greatly assists drivers, travellers or public security and safety in their decision-making. In particular, Spatio-Temporal Graph Neural Networks (STGNNs) have gained popularity as a powerful tool for effectively modelling spatio-temporal dependencies in diverse real-world urban applications, including intelligent transportation and public safety. However, the black-box nature of these models prevents a true understanding of the results and a trustworthy adoption by their users. The research field of eXplainable Artificial Intelligence (XAI) addresses this concern by developing systems that help users trust non-transparent AI. Non-expert ITS users are primarily interested in the non-technical reasons behind model predictions. Hence, leveraging Natural Language Generation (NLG), verbal descriptions of the reasons behind model outcomes are a peculiar tool to provide an easy and clear illustration of the process. This work focuses on developing an XAI system to explain short-term speed forecasts in traffic networks obtained from STGNNs. The primary emphasis lies in explaining the reasons behind predicted traffic congestions or free flows. Key information justifying these predictions is extracted from the input traffic network in the form of a significant subgraph. The information of the subgraph is finally summarized and it is then converted into text using a template-based approach. This thesis makes a dual contribution. First, it delves into explaining short-term traffic predictions by STGNNs, an underexplored area in current literature. Second, it leverages NLG techniques uniquely, using them to verbally translate XAI explanations into a coherent narrative, following a data-to-sequence template-based approach.