Interpretable Traffic Flow Prediction Using SHAP and STID-Recurrence Model

Efficient traffic flow prediction is critical for reducing congestion, lowering emissions, and improving urban mobility in smart cities. However, the lack of interpretability in complex AI models limits their practical adoption for decision-making. This paper addresses the challenge of interpreting spatiotemporal models by applying the STID-recurrence model to traffic flow forecasting on the PEMS-BAY dataset. We use SHapley Additive exPlanations (SHAP) to study the contribution of individual traffic sensors to model predictions, identifying significant sensors around major highways and intersections. Temporal analysis reveals how sensor importance changes over peak and off-peak hours and across different days of the week, reflecting realistic traffic dynamics. Experimental results show that discarding low-influence sensors does not significantly affect predictive accuracy, suggesting opportunities for cost-effective sensor optimization, this study is limited to the PEMS-BAY dataset, and the generalizability of findings to other urban environments or datasets remains to be validated, despite These findings, the results show enhancement of the interpretability of spatiotemporal models, providing actionable insights for traffic management in smart cities and paving the way for scalable, data-driven solutions.