category_student_information

MSc thesis topic: Explainable Artificial Intelligence in Mobility Prediction

Artificial intelligence (AI) is revolutionizing many applications in Geo-Information Science. The transportation sector offers many opportunities to apply the progress in AI for intelligent transportation systems. Such intelligent transportation systems must be built from an intricate combination of artificial intelligence, computing & data infrastructure and mobility analysis.

Artificial intelligence (AI) is revolutionizing many applications in Geo-Information Science. The transportation sector offers many opportunities to apply the progress in AI for intelligent transportation systems. Such intelligent transportation systems must be built from an intricate combination of artificial intelligence, computing & data infrastructure and mobility analysis.

In transportation applications deploying advanced deep neural networks has been slow because such deep neural networks are difficult to interpret and lack robustness. Increasingly, research efforts try to address this by developing interpretable and robust machine learning methods, such as causal inference approaches that provides interpretable and actionable information. At the same time, running deep learning algorithms requires the setup of complex software dependencies and is difficult to operate for non-expert. Digital twins offer a one-stop solution to host explainable AI methods and expose their parameters to users. On the side of mobility analysis, there is also a deficit with most of these explainable AI methods as they are developed for image or sequential data which is insufficient to address unique challenges in mobility data analysis.

This thesis contributes to making AI for mobility more interpretable, reliable, and reusable. Thus, contributing to safer, more efficient, and more sustainable future for mobility systems:

  • The thesis uses the Interpretable and Robust Machine Learning for Mobility Analysis (IRMLMA) framework to assess the robustness of mobility predictions with causal intervention. This framework offers direct control of behavioral dynamics, emerging mobility patterns, and evaluating the network performance.
  • The framework is straightforward in its utilization and flexible customization of its components due to the integration into the Open Digital Twin Platform (ODTP).

Predicting individual mobility accurately is required to enable mobility services (Ma and Zhang, 2022) and serves as the backend for intelligent transport systems (Tang et al., 2019). Customer satisfaction and system efficiency can be improved with proactively offering predictions of individual mobility (Zhao et al., 2018). In turn, this helps in decarbonizing the transport sector (Hong et al., 2022). Individual mobility prediction remains a challenging problem due to the complexity in mobility patterns which are influenced by a wide range of behavioral factors and spatial contexts (Hong et al., 2023). These spatiotemporal dependencies generally hinder predictions of individual mobility (Wiedemann et al., 2023a).

In recent years, the availability of human movement traces and the advancements in data-driven models have significantly enhanced mobility prediction ability (Luca et al., 2021). Modern neural networks reached acceptable performance but are criticised for their low interpretability (Pappalardo et al., 2023). In the context of the decision-making, these networks are commonly regarded as “black boxes” because it is nearly impossible to reconstruct the reasoning behind a prediction. For mobility prediction, the lack of interpretability makes it impossible to judge which spatiotemporal patterns are represented and what role behavioral factors played. So far, these models hinder decision-making and policy design, and are considered unreliability and untrustworthiness by the practitioners (Huang et al., 2020; Koushik et al., 2020). The multitude of available data contrasts with the scarcity of publicly available individual mobility data sets due to privacy concerns (Wiedemann et al., 2023b). Without comparable data comparing mobility prediction models (Graser et al., 2023) becomes impossible.

Causal intervention offers a novel tool to process and generate data from diverse environments. It also enables the assessment of neural network robustness and providing human-readable causal inference for interventions (Xin et al., 2022). Building upon the advantages of this approach, the student will use a framework for systematically evaluating the impact of mobility behaviors on the performance of deep learning predictions called IRMLMA. This framework also quantifies uncertainty and distributional robustness to provide confidence levels that are essential for reliable decision-making.

Complex simulations and machine-learning models such as IRMLMA increase in application in research, industry, and governance. However, applying these systems with reasonable accuracy and efficiency requires large-scale efforts of data collection, data transformation, data analysis, and data visualization. At the same time, maintaining the required infrastructure, software, and personnel skyrockets, making these tools unavailable to many potential users. The paradigm of the digital twin offers a novel perspective on how to manage the data efficiently and make these systems available more steadily at a lower cost. The Open Digital Twin Platform (ODTP) is designed to be openly available to all interested parties to enable a common framework and baseline for digital twin based research. ODTP uses containerization, loose coupling, and micro-services to provide dynamically composable digital twins. ODTP also provides tools for licensing resolution, privacy and access control, and reproducibility. For this thesis, ODTP implements a common mobility research pipeline for IRMLMA. These programs are often difficult to assemble and use, thus leading to dangerous versions of “never change a running system”. ODTP converts them into an easy-to-use version, making it possible to initiate mobility simulations with one click. ODTP enables the quick adding of relevant data sources and analytical pipelines related to any topic and make them easily usable, accessible, and shareable to research, industry, and governance.

Relevance to research/projects at GRS or other groups

  • One of the strategic initiatives of WUR is the creation and exploration of Digital Twins. GRS is leading efforts to create digital twins for mobility applications. This project is in cooperation with TU Delft and the scientific lead of the XAI framework IRMLMA used in this thesis.

Objectives and Research questions

Researching explainable AI (XAI) offers a broad range of tasks. Master students interested in the following topics can approach me for further details:

  • Working with Dutch trajectory data
  • Exploring causal inference XAI approach for the 30km city in Amsterdam
  • Experimenting on neural networks’ robustness against domain shifts

Requirements

  • Good python knowledge (required)
  • Machine learning and deep learning basics (required)
  • Causal inference (optional)
  • Digital Twin background (optional)
  • Transport & Mobility research interest (optional)

Mandatory reading

Recommended reading

  • Dirmeier, S., Hong, Y., Xin, Y., & Perez-Cruz, F. (2023). Uncertainty quantification and out-of-distribution detection using surjective normalizing flows. arXiv preprint arXiv:2311.00377.
  • Ma, Z., & Zhang, P. (2023) Individual mobility prediction review: Data, problem, method and application. Multimodal Transportation, 1(1):100002. doi:10.1016/j.multra.2022.100002.
  • Tang, Y., Cheng, N., Wu, W., Wang, M., Dai, Y., & Shen, X. (2019). Delay-Minimization Routing for Heterogeneous VANETs With Machine Learning Based Mobility Prediction. IEEE Transactions on Vehicular Technology, 68(4):3967–3979. doi:10.1109/TVT.2019.2899627.
  • Zhao, Z., Koutsopoulos, H. N. , & Zhao, J. (2018). Individual mobility prediction using transit smart card data. Transportation Research Part C: Emerging Technologies, 89:19–34, 2018. doi:10.1016/j.trc.2018.01.022.
  • Hong, Y., Martin, H., & Raubal, M. (2022). How do you go where? improving next location prediction by learning travel mode information using transformers. In Proceedings of the 30th International Conference on Advances in Geographic Information Systems (SIGSPATIAL ’22). doi:10.1145/3557915.3560996.
  • Hong, Y., Zhang, Y., Schindler, K. & Raubal., M. (2023). Context-aware multi-head self-attentional neural network model for next location prediction. Transportation Research Part C: Emerging Technologies, 156:104315, 2023b. doi:10.1016/j.trc.2023.104315.
  • Wiedemann, N., Hong, Y., & Raubal, M. (2023a). Predicting visit frequencies to new places. In 12th International Conference on Geographic Information Science (GIScience ’23), volume 277, pages 84:1–84:6. doi:10.4230/LIPIcs.GIScience.2023.84.
  • Luca, M., Barlacchi, G., Lepri, B., & Pappalardo, L. (2021). A Survey on Deep Learning for Human Mobility. ACM Computing Surveys, 55:7:1–7:44, doi:10.1145/3485125.
  • Pappalardo, L., Manley, E., Sekara, V., & Alessandretti, L. (2023). Future directions in human mobility science. Nature Computational Science, 3(7):588–600. doi:10.1038/s43588-023-00469-4.
  • Huang, X., Kroening, D., Ruan, W., Sharp, J., Sun, Y., Thamo, E., Wu, M., & Yi,X. (2020). A survey of safety and trustworthiness of deep neural networks: Verification, testing, adversarial attack and defence, and interpretability. Computer Science Review, 37:100270.
  • Koushik, A. N., Manoj, M., & Nezamuddin, N. (2020). Machine learning applications in activity-travel behaviour research: a review. Transport Reviews, 40(3):288–311. doi:10.1080/01441647.2019.1704307.
  • Wiedemann, N., Martin, H., Suel, E., Hong, Y., & Xin, Y. (2023b). Influence of tracking duration on the privacy of individual mobility graphs. Journal of Location Based Services, 0(0):1–19. doi:10.1080/17489725.2023.2239190.
  • Graser, A., Jalali, A., Lampert, J., Weißenfeld, A., & Janowicz, K. (2023). Deep Learning From Trajectory Data: a Review of Deep Neural Networks and the Trajectory Data Representations to Train Them. In Proceedings of the Workshop on Big Mobility Data Analytics (BMDA) co-located with EDBT/ICDT 2023 Joint Conference.

Theme(s): Modelling & visualisation, Human – space interaction