An Explainable Spatio-Temporal Framework for Traffic Forecasting Using Graph-Based Features

Abstract

Urban traffic prediction plays a critical role in improving transportation efficiency and supporting sustainability in smart cities. This study proposes an explainable spatio-temporal machine learning framework that integrates periodic temporal features with graph-based spatial representations. Temporal dynamics are modeled using cyclical transformations of time-related variables, while spatial dependencies are captured through graph centrality metrics derived from the transportation network. The proposed approach is evaluated on real-world traffic data collected from seven critical intersections in Sakarya, Türkiye. To ensure a comprehensive assessment, multiple models including Random Forest, XGBoost, and Support Vector Regression (SVR) are compared. Experimental results show that tree-based ensemble methods perform significantly better, and the highest prediction performance is consistently obtained with the spatio-temporal feature configuration ( $R^{2} \approx 0.96$ ). These findings demonstrate that traffic flow is not only governed by temporal patterns but is also strongly influenced by spatial interactions within the network. In addition, the impact of lag-based features, which are widely used in the literature, is systematically analyzed. Results indicate that although lag variables appear highly important in feature importance analyses, their contribution to predictive performance is limited. This highlights a critical distinction between feature importance and actual predictive utility, referred to in this study as misleading importance. Overall, the proposed approach provides a computationally efficient and interpretable solution for urban traffic prediction, while also providing insights into feature selection in spatio-temporal modeling.