Interpretable machine learning for traffic congestion prediction: Unveiling the impact of different COVID-19 periods

Traffic congestion prediction plays a crucial role in mitigating congestion. However, the COVID-19 pandemic and associated government control measures have significantly altered urban travel behavior, increasing the complexity of traffic congestion prediction. This study aims to predict traffic congestion in Alameda County in the San Francisco Bay Area, USA, during the prelockdown, lockdown, and postlockdown periods. We incorporate three external categories of data, i.e., weather conditions, seasonality factors, and COVID-19-related variables, and use recursive feature elimination with cross-validation to identify important features across different periods and avoid potential overfitting. On this basis, multiple advanced machine learning (ML) models, including support vector regression (SVR), multiple linear regression (MLR), recurrent neural network (RNN), and long short-term memory (LSTM) networks, are trained and optimized through extensive experimentation and parameter tuning. Since LSTM has more hyperparameters and is more sensitive to tuning than the other ML methods used, we employ an adaptive parameter selection approach to optimize its hyperparameters, enhancing model accuracy and efficiency, rather than manually tuning parameters for SVR and RNN. These models are evaluated via the normalized root mean square error. The results indicate that the bidirectional LSTM (Bi-LSTM) consistently outperforms the other models across all COVID-19 periods. This superior performance can be attributed to the Bi-LSTM's bidirectional architecture, which effectively captures temporal dependencies by analyzing data both forward and backward in time. To address the limited interpretability of ML methods and provide valuable insights, we apply the integrated gradient (IG) technique to interpret the best-performing and differentiable Bi-LSTM predictions. Our analysis revealed that new COVID-19 cases had a negative influence on traffic congestion during the lockdown and postlockdown periods. The observed reduction in traffic can be explained by heightened public risk awareness, voluntary reductions in travel, and compliance with government-imposed mobility restrictions. We also apply SHapley Additive exPlanations to SVR, given that IG is not applicable to this model. The results indicate that in the postpandemic period, people have become more cautious—high new hospitalization discourages travel, reducing traffic congestion, whereas high fuel prices do not deter a shift toward private vehicle use, leading to increased congestion.