信控交叉口电动自行车违规行为已成为城市交通系统中日益严重的安全隐患，极大地威胁交通秩序与安全。为了实现对违规行为的精细研判与动态干预，本文提出了一种结合生存分析与机器学习的电动自行车违规行为分析方法，揭示违规行为的时序演化规律及其关键驱动因素。首先结合Kaplan-Meier方法与Cox比例风险模型，建立了电动自行车违规风险函数，分析电动自行车违规行为的动态演化规律。其次引入基于梯度提升框架的轻量集成学习模型LightGBM，建立了电动自行车违规行为研判方法，并采用SHAP分析提升了模型的可解释性，研判电动自行车在信控交叉口违规行为，探究多维度环境与个体特征对违规风险的综合影响。最后，基于视频拍摄提取的576条有效样本进行了多交叉口实证研究。电动自行车违规风险函数结果表明，二次过街、右转冲突和外卖骑手身份显著增加了违规概率。载人骑行者的违规风险为不载人骑行者的85%，让行行为能够减少约54%的违规风险。本文提出的可解释LightGBM模型违规行为研判准确率更高，且在训练速度上远超随机森林等模型，表明了其在电动自行车违规风险研判方面的优越性。模型结果表明：前车跟随、观察左右来车、二次过街和组群规模是影响违规行为的核心要素。当骑行者处于3人以上的组群，个体的责任感减弱，产生“法不责众”的心理，显著增加了违规风险。研究成果可为电动自行车违规行为的精细化治理以及制定更为精准的交通政策提供理论依据。

Violation behaviors of electric bicycles at signal-controlled intersections have become an increasingly serious safety hazard in urban traffic systems, posing a significant threat to traffic order and safety. To achieve refined assessment and dynamic intervention of such violations, this study proposes an analytical method that integrates survival analysis with machine learning to reveal the temporal evolution patterns and key driving factors of electric bicycle violation behaviors. First, by combining the Kaplan–Meier method and the Cox proportional hazards model, a violation rate function for electric bicycles is established to analyze the temporal characteristics of violation behaviors. Second, a lightweight ensemble learning model, LightGBM, based on the gradient boosting framework, is introduced to develop a violation behavior inference approach. The SHAP analysis is further employed to enhance model interpretability, enabling a comprehensive examination of electric bicycle violation behaviors at signalized intersections and exploring the combined effects of multidimensional environmental and individual characteristics on violation risk. Finally, an empirical study based on nearly 600 samples extracted from video recordings is conducted. The results of the electric bicycle violation rate function indicate that secondary crossing, right-turn conflicts, and rider identity significantly increase the probability of violations. Riders carrying passengers exhibit a lower probability of violation, with a violation risk equal to 85% of that of non-passenger riders. Yielding behavior markedly reduces the probability of violation, lowering the violation risk by approximately 54%. Owing to the gradient boosting optimization strategy, the interpretable LightGBM model proposed in this study achieves higher accuracy in violation behavior inference and significantly faster training speed than models such as random forests, demonstrating its superiority in assessing the violation risks of electric bicycles. The model results further show that following behavior, observation of cross traffic, secondary crossing, and group size are the core factors influencing violation behaviors. Following behavior triggers a “herd effect,” and an increase in group size strengthens the sense of “deindividuation,” leading to a significant rise in violation probability as the group expands. When riders travel in groups of more than three, individual responsibility weakens, and a “law does not punish the crowd” mentality emerges, substantially increasing the violation risk. The findings provide theoretical support for the refined governance of electric bicycle violation behaviors and the development of more precise traffic management policies.