Academic Awards 2025 booklet

53 Leveraging Contrastive Learning and Spatial Encoding for Traffic Network Prediction with Expanding Infrastructure Urban transportation networks are evolving rapidly, driven by expanding infrastructure and the growing complexity of city life. As cities install and relocate traffic sensors to manage congestion and optimize traffic flow, a critical challenge emerges: how can we accurately predict traffic conditions for newly added sensors that existing models have never seen? My research addresses this gap by introducing the Distance- Enhanced Spatial Pre-Training (DESPT) framework—a novel approach that fuses spatial encoding with contrastive learning to dramatically improve real-time traffic predictions. DESPT stands out by integrating routing-based spatial relationships and advanced pre-training techniques, enabling models to “learn” from both historical data and the dynamic structure of sensor networks. Through rigorous testing on real-world datasets from Los Angeles, the San Francisco Bay Area, and The Hague, DESPT demonstrated significant improvements in predictive accuracy for newly deployed sensors. By bridging the worlds of artificial intelligence and urban planning, this work not only pushes the boundaries of traffic forecasting but also lays a foundation for smarter, more adaptive cities. The innovations presented here promise more efficient, responsive urban mobility, with direct benefits for city planners, commuters, and the communities they serve. This research introduces an innovative framework combining spatial encoding and contrastive learning to improve traffic prediction, especially for new sensors in urban networks. Tested on real-world datasets, the model significantly boosts prediction accuracy, paving the way for smarter, more adaptive cities and offering direct benefits to urban mobility management. Figure 1: Distance-Enhanced Spatial Pre-Training framework. Figure 2: Spatial Encoding framework.