A Graph Neural Network-based Approach to Predict the Effects of Urban Climate on Personal Mobilty Choices in Seoul, South Korea
M. Giampaolo (TU Delft - Architecture and the Built Environment)
A. Rafiee – Mentor (TU Delft - Architecture and the Built Environment)
M.A. Mosteiro Romero – Mentor (TU Delft - Architecture and the Built Environment)
S. Rahmani – Graduation committee member (TU Delft - Civil Engineering & Geosciences)
More Info
expand_more
Other than for strictly personal use, it is not permitted to download, forward or distribute the text or part of it, without the consent of the author(s) and/or copyright holder(s), unless the work is under an open content license such as Creative Commons.
Abstract
Urban climate affects how people move through cities, but its influence is difficult to capture with models based on generalized comfort indices that ignore individual experiences of climate. This thesis instead explores a bottom-up approach that uses daily Global Navigation Satellite System (GNSS) traces of people traversing an urban environment, which inherently contain each individual's personal influences on their mobility. A machine learning model was developed and trained using this dataset with the purpose of predicting future mobility values, while assessing the role that climate played in such predictions. The model employed a Spatio-Temporal Graph Neural Network (STGNN) architecture to capture both potential spatial dependencies between visited locations and temporal patterns in their activity.
The work draws on the Seoul Cozie dataset, which recorded six weeks of GNSS location data from wearable devices of 22 participants in Seoul during autumn 2023. Positions were aggregated into a graph structure with road intersections as nodes and transitions between them as edges. Climate features (temperature, humidity and PM10) were interpolated from over 1,000 weather stations using a Triangulated Irregular Network method and added as dynamic node features. STGNN variants were trained and compared based on whether they included climate node features.
Results show forecasts of node visits with low Mean Square Error of around 0.12 . However, precision and recall values for visited/unvisited node detection are low, peaking at 56.51%, reflecting strong class imbalance in the input. Adding climate attributes produced only minor and inconclusive improvements, in part due to the dataset’s short time span. The thesis proposes a reproducible framework linking climate and mobility, while underlining the need for richer datasets and for more flexible model architectures, capable of addressing class imbalances and representing personal mobility datasets.