驾驶风格是驾驶行为的外在表达，激进风格的驾驶员容易进行更为频繁的危险驾驶操作，加剧车辆之间的交互作用，影响换道安全。在换道动作执行前识别驾驶员的驾驶风格，可以通过个性化预警信息有效约束驾驶员行为。文中提出了一种网联环境下考虑换道博弈的驾驶风格在线识别方法SHAP-XGBoost，以期在换道意图期间完成驾驶风格的识别。首先，将换道意图期间换道车辆及其周围车辆的个体行为和博弈行为的波动程度作为输入特征变量，通过相关性分析、主成分分析以及4种不同聚类方法对驾驶风格进行标记；然后，利用提出的SHAP-XGBoost模型选择关键特征，以训练驾驶风格识别模型，并通过滑动窗口完成在线识别；最后，采用HighD数据进行实验。结果表明：与基于质心距离、连通性、密度分布的聚类方法相比，基于图论原理的谱聚类可以更好地根据输入特征变量的形态标记驾驶风格；利用SHAP-XGBoost模型及14个关键特征进行驾驶风格识别，可以在不损失准确率的同时提高在线识别效率，驾驶风格识别准确率高达99%；同时将个体特征和博弈特征作为模型的输入时，可以提升驾驶风格标记和识别的准确率。此研究成果可为个性化换道决策和预警提供支持。

Driving style is the external expression of driving behavior. Drivers with aggressive style tend to engage in more frequent risky driving operations, intensifying interactions between vehicles and affecting lane-changing safety. Identifying a driver’s driving style before executing a lane-changing can effectively constrain driver’s behavior through personalized warning information. This paper proposed the SHAP-XGBoost method, which considers lane-changing game in a connected environment, aiming to achieve the real-time recognition of driving styles during the lane-changing intention phase. Firstly, the fluctuation degree of individual behavior and gaming behavior during the lane-changing intention was used as input feature variables, and the driving style was marked by correlation analysis, principal component analysis, and four different clustering methods. Next, the proposed SHAP-XGBoost model was used to select key features for training the driving style recognition model, and online recognition was completed through a sliding window. Finally, experiments were conducted using the HighD dataset. Results show that: compared with clustering methods based on centroid distance, connectivity and density distribution, spectral clustering based on graph theory principles can better label driving styles based on the morphology of the input feature variables; using the proposed SHAP-XGBoost model with 14 key features for driving style recognition can improve online recognition efficiency without loss of accuracy, and the driving style recognition accuracy is up to 99%; simultaneously incorporating individual features and gaming features as inputs to the model can improve the accuracy of driving style labeling and recognition. The research results can be used to support personalized lane-changing decisions and early warnings.