Multi-Task Learning for Ski Injury Predictions

Abstract

Predicting ski injuries is a very hard classification problem. This is due to the high class imbalance of injured vs. non-injured skiers and the lack of demographic information about skiers. Additional problems are the intrinsic properties of the ski lifts. Ski lifts differ in width, the difficulty degree, geographical position on the mountain etc. which results in different patterns for ski injury. In most researches, this information is not included. Aim of this paper is to develop multi-task classification models, which account for the uniqueness of ski lifts, taking into consideration information from other ski lifts. The proposed models were created on Mt. Kopaonik, Serbia ski resort and they show that ski injury in the following hour can be predicted with AUC similar to 0.64, or 3-4% better compared to the classical approaches.