Analyzing Disparities in Machine Learning-based Ski Injury Prediction: A Longitudinal Study for Mt. Kopaonik Ski Resort

Abstract

Using machine learning algorithms to predict the occurrence of ski injuries in recreational skiing is not a novel task. Although the number of injuries during skiing is generally very low, the cost of injuries can be very high. The goal of this paper is to analyse disparities and error discrepancies within a prediction model for a ski resort learned on multiple ski regions and different periods of the day. The data originates from Mt. Kopaonik, Serbia, a resort with around 1.5 injuries per thousand skier days, and comprises six consecutive seasons of ski lift gate entrances aggregated to an hourly level per ski region. The goal of the prediction was to predict whether an injury will occur on the ski slope within the following hour, training on one season and applying the model to the following one. The results indicate a slight decline in AUC over six seasons, with the most popular ski regions maintaining stable AUCs close to the overall value. Ski slopes with fewer skiers fall into distinct clusters: easy-to-predict regions of lower difficulty, difficult-to-predict regions where beginners learn to ski, and unstable regions whose AUCs oscillate strongly across years due to frequently changing conditions. The best predictive performance occurs in the morning hours, when fewer skiers are on the slopes, while the worst is observed in the afternoon.

Sandro Radovanović

Assistant Professor at University of Belgrade

My research interests include machine learning, development and design of decision support systems, decision theory, and fairness and justice concepts in algorithmic decision making.