Traffic prediction is an active research area to monitor and manage the evolving demand and supply dynamics. Previous research has established that deep learning, specifically Graph Neural Networks (GNNs), can help to obtain accurate predictions. However, much focus is on optimizing the model accuracy using pre-processed and standardized datasets. Transfer learning offers potential by using pre-trained traffic prediction models for new data or study areas. However, a few challenges also arise while transferring the pre-trained models for practical applications since real-world data are anything “ideal.” Further, the interpretability or explainability of the GNN models is not usually explored in the context of traffic prediction. In the MINDMAP project, we research and test the theoretical and practical methods for data-efficient and interpretable transfer of the traffic prediction models. On the one hand, data-centric methods will provide systematic approaches to transforming the raw data into learnable representations. On the other hand, interpretability methods will provide insights and transparency into the working of the GNNs during transfer by highlighting which elements of demand and supply are transferable and vice-versa. Finally, the developed approaches will be integrated into a prototype/dashboard for real-time network-wide traffic prediction. Overall, MINDMAP will not only extend the state-of-the-art in traffic prediction but will also help to bridge the gap between research and practice.

The three groups collaborating on this international project are the chair of Transportation Systems Engineering (TSE) of TUM (Prof. Dr. Constantinos Antoniou), the group of Prof. Dr. Jiwon Kim from the University of Queensland, Australia, and the Data Analytics and Machine Learning (DAML) group of TUM (Prof. Dr. Stephan Günnemann). The research team will approach the ambitious research goals set out in this proposal from the perspectives of transferability, interpretability, and data-centric perspectives, using distinct but complementary methodologies and access to different sets of collaborators and resources while striving for a common goal.