Bridging Efficient and Explainable Traffic Flow Prediction on the Edge

Digital transformation is significantly reshaping the way transportation systems operate, bringing both new opportunities and challenges. Indeed, Intelligent Transportation Systems (ITSs) are facing accurate traffic flow prediction as a fundamental means for congestion mitigation, route planning, and dynamic traffic signal control, by leveraging machine learning techniques. While deep learning models, like Recurrent Neural Networks, have proved strong predictive performance, they require substantial computational resources and lack interpretability, hindering back their adoption on low-end edge devices and failing to provide required robusteness of real-world applications, as required by regulatory standards. Bearing in mind the above, in this paper, we devise an approach that combines local execution onto edge devices of traffic flow prediction by using XGBoost, a tree-ensemble machine learning model, with SHapley Additive exPlanations (SHAP) technique, which guarantees explainability. We detail the whole flow and prove its effectiveness by running it aganist traffic dataset, showing that we are able to achieve competitive accuracy and local-interpretation of predictions.