Experimental and interpretable machine learning-based analysis of pedestrian evacuation behavior in attack situations

Previous studies have extensively examined pedestrian evacuation behavior during emergencies such as fires. However, few studies have focused on evacuation behavior during violent attacks. In this paper, we design an experiment to simulate a violent attack and study the evacuation behavior of pedestrians during such an event. A random forest model is used to predict evacuation outcomes based on experimental data. Additionally, we employ interpretable machine learning methods, specifically Partial Dependence Plots (PDPs) and Individual Conditional Expectation (ICE), to investigate the variables influencing evacuation outcomes. The results indicate that the distance variable is the key variable influencing evacuation outcomes, followed by preparation time. Generally, shorter preparation time is beneficial to a higher probability of successful evacuation. However, immediate action after identification of the attacker is instead associated with a lower probability of successful evacuation. Moreover, shorter preparation time is also related to lower probabilities of successful evacuation when approaching the attacker. The number of attackers and exits has exerted a relatively limited monotonic influence. This study contributes to the development of safety guidelines and contingency plans in the event of a violent attack.