Exploring the Efficacy of Artificial Intelligence in Speed Prediction: Explainable Machine-Learning Approach

The primary concern regarding road design elements is traffic stream speed, which is considered a good indicator of travel quality. Although a great deal of time and effort is needed to estimate traffic stream speed from the field, it still frequently falls short of the designers’ requirements. To address this, this study explored the efficacy of artificial intelligence, and used eXplainable Machine Learning (XML) to determine the underlying mechanisms through which machine-learning models arrive at predictions. This study investigated the impact of categorized traffic volume and road width on predicting stream and vehicle-specific speeds, considering five significant vehicle categories in India’s traffic stream. Because small cars and two-wheelers account for the highest proportion of the five vehicle categories, speed prediction models for these vehicles, along with stream speed, are proposed. Four tree-based machine-learning models were used: decision tree (DT), random forest (RF), extra tree (ET), and eXtreme Gradient Boosting (XGB). The performance of all the models was validated, and the outcomes showed that the RF had the best predicting speed. Furthermore, a test data set was used as an independent and unseen sample used to assess the final performance of the developed model fits impartially. A post hoc explanation technique, i.e., SHapley Additive exPlanations (SHAP), was employed for the complex RF model to interpret. SHAP indicated various positive and negative impacts of categorized traffic volume and road width on predicting stream speed, illustrating these relationships and aligning with established traffic engineering principles. This analysis with SHAP validated the causal relationship behind the ML model’s predictions. The study outcomes are useful for managing the urban roadway network performance under mixed traffic conditions.