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

•New information on evacuation behavior is provided to reduce casualties in violent attacks in classrooms.•Using interpretable machine learning methods to investigate behavior of evacuees during violent attacks in classrooms. We build machine learning models to predict evacuation outcomes and then use the SHAP tool to explain the predictions.•We fully considered the particular features of the classroom layout and took the attacker's attack route into consideration, unlike the experimental design for evacuation of ordinary buildings.