NAS for Traffic Prediction: A Survey of Methods and Challenges.

Traffic prediction is a critical component of intelligent transportation systems, supporting applications from adaptive signal control to route guidance. While deep learning models have achieved strong results, their architectures are typically designed manually, demanding significant expert effort and often struggling to generalize across different cities or datasets. This survey explores Neural Architecture Search (NAS) as a systematic alternative to manual design for traffic prediction models. NAS automates the process of finding optimal deep learning architectures that effectively capture the complex spatial-temporal patterns inherent in traffic data, eliminating trial and error. The paper categorizes NAS methods by search strategy (gradient-based, evolutionary, one-shot weight-sharing), detailing how search spaces are designed for spatial and temporal traffic operators and how strategies balance cost against architecture quality. It also highlights key challenges, including computational scalability for large road networks, the need for better cross-city generalization, handling dynamic graph structures, and the potential for NAS in developing spatial-temporal foundation models.