SAFER-predictor: Sparse adversarial training framework for robust traffic prediction under missing and noisy data

Accurate traffic flow forecasting is essential for developing intelligent transportation systems (ITSs) to reduce congestion, optimize road management, and improve safety. While data-driven traffic prediction approaches have shown high accuracy, they rely heavily on precise measurements, making them vulnerable to perturbed environmental factors, like sensor malfunctions, data storage issues, and adverse weather conditions. To overcome the limitation, we propose SAFER-Predictor, a novel sparse adversarial training (Sparse AT) framework for enhancing the reliability of deep learning based spatiotemporal traffic prediction models. Sparse AT extends traditional adversarial training (AT) through a two-phase process: pre-training and fine-tuning. In the pre-training phase, the model is optimized to capture normal traffic patterns, enhancing predictive performance by understanding standard dynamics without external disruptions. In the fine-tuning phase, the focus shifts to strengthening robustness against corrupted inputs by employing an iterative min-max strategy during AT, optimizing performance for worst-case scenarios. Furthermore, we derive theoretical formulations that establish an upper bound on the model's prediction error following Sparse AT under certain noise levels. Experimental results indicate that incorporating Sparse AT into the representative traffic flow prediction models improves stability and ensures high accuracy under various perturbation scenarios.