Evacuation behavior under violent attacks in classrooms based on experiments and interpretable machine learning method

Abstract

Extremely violent crimes seriously affect campus safety. However, there is a lack of studies on evacuation behavior in such events, and there are no effective strategies for emergency response and evacuation. To better understand evacuation behavior in campus violent attacks, this paper designs an experiment of classroom violent attacks and tries to predict the evacuation outcome with random forest algorithm based on the experimental data. Then an interpretable machine learning method named Shapley Addictive Explanations is used to study the factors affecting the evacuation outcome. The results show that the random forest model outperforms other models in prediction, with an accuracy of 96.5% on the validation set. The mean distance between participants and attacker is the most important factor with a positive effect. The second most important variable is the evacuation preparation time, with shorter preparation time being related to a higher probability of successful evacuation, except in cases where the attacker is noticed just before the attack is about to occur. Participants are more likely to successfully evacuate if they are seated in the middle row rather than in the front and back, but there exists a complex interaction between the initial location and attack route. The attack route and classroom type have a small influence, while the effect is monotonic: the attacker entering through the front door, as well as the examination classroom layout, are associated with increased chances of a successful evacuation. This study can contribute to developing security guidelines and contingency plans for campus violent attacks.