A two-stage driving-style-stratified framework for analyzing and predicting car-following conflict risk in mixed traffic.

The growing coexistence of automated vehicles (AVs) and human-driven vehicles (HVs) in mixed traffic fundamentally alters car-following dynamics and elevates rear-end conflict risk. However, human drivers exhibit marked heterogeneity in driving styles-conservative, normal, or aggressive-which significantly moderates how kinematic factors translate into conflict risk. Ignoring such style‑dependent heterogeneity leads to biased risk analysis and ineffective countermeasures. Moreover, existing conflict prediction models typically output deterministic point estimates without quantifying predictive uncertainty, a critical shortcoming for safety‑critical automated driving decisions. To address these gaps, this study proposes a two‑stage, driving‑style‑stratified framework that jointly analyzes conflict risk mechanisms and delivers uncertainty‑aware predictions. Using the Lyft Level-5 autonomous driving dataset, we extract microscopic driving volatility indicators, quantify conflict risk via the Rear-end Conflict Risk Index (RCRI), and identify three driving styles through K-medoids clustering. In Stage 1, a Random Parameters Logit model with Heterogeneity in Means and Variances (RPLHMV) uncovers key risk factors and their multi-layered heterogeneity across AV‑following‑HV (AV‑HV) and HV‑following‑AV (HV‑AV) scenarios and driving styles. In Stage 2, informed by these insights, a Random Deep & Cross Network with Monte Carlo dropout (RDCN-MC) explicitly models behavioral heterogeneity and quantifies predictive uncertainty. Results show the significance and marginal effects of risk factors vary markedly across driving styles, with aggressive drivers exhibiting the strongest influences. The RDCN-MC model achieves accuracy, precision, recall, and F1-score above 97 %, 87 %, 93 %, and 90 % under class imbalance, consistently outperforming benchmarks. Each prediction is accompanied by an uncertainty estimate that effectively discriminates correct from incorrect predictions, serving as a confidence signal. Based on these findings, we propose differentiated, style-specific safety strategies for trustworthy automated driving control and human driving, thereby supporting proactive safety management in mixed traffic.