A Machine Learning-Based Framework for Risk Recognition and Reliability Evaluation in City Expressway Ramp Merging

Risk identification in ramp merging is often compromised by complex vehicle interactions and ‘future information’ leakage. To resolve this, we decouple the process using an ‘observation-conflict’ mechanism. By extracting kinematic features solely from the merging preparation phase, the framework predicts continuous risks during actual execution without temporal bias. Furthermore, we stabilize risk detection by integrating cpTTC thresholds with duration constraints into a three-level risk labeling scheme, ensuring the results align with real-world physical dynamics. Multiple machine learning models are comparatively evaluated using group-based data partitioning. Results indicate that the XGBoost model achieves the best overall performance, yielding an overall accuracy of 0.8182 and a multiclass AUC (OvR) of 0.8898. Furthermore, time cross-domain validation under varying macroscopic traffic flow states demonstrates that the framework exhibits reliable statistical stability; when reconstructed into a binary classification task, it maintains a risk recall of 0.9978. These findings provide a reliable methodological basis and early-warning feature reference for dynamic traffic risk assessment in merging scenarios.