Urban Topology Foundation Models for Sample-Efficient Urban Traffic Imputation

Kurzbeschreibung

Beschreibung Data-driven traffic state estimation pipelines typically rely on tabular models that operate in a topologically blind manner, requiring high sensor density to map spatial features to flow. This thesis develops the Urban Topology Foundation Model (UTFM), a graph-based representation learning framework designed to achieve extreme sample efficiency in traffic imputation. By pre-training a Graph Neural Network architecture on global urban geometries via Self-Supervised Learning (SSL), the model extracts universal latent structural embeddings of road networks. The primary objective is to evaluate the downstream predictive capacity of these embeddings combined with a differentiable prediction head, mapping the imputation performance frontier across sparse sensor coverage regimes. Voraussetzungen • Programming Proficiency: Advanced proficiency in Python programming and geometric deep learning libraries, specifically PyTorch. • Machine Learning Foundations: Solid theoretical and practical understanding of Neural Networks, ideally experience with GNNs, Transformers, and Self-Supervised Learning (SSL) paradigms.