Can ensemble learning approaches for offside detection work?

Abstract

The analysis of data collected from various recreational activities and professional sports is essential to obtain more information on the activity in question or to make better data-driven decisions. Most literature related to offside detection related to the efficacy of manual offside detection or the use of an offside detection algorithm. In this study, the focus shall be on the detection of offside judgements in football/soccer using ensemble learning approaches such as random forest type algorithms, boosting type algorithms and majority voting. For random forests, we also consider three corresponding extensions: regularized random forests, guided regularized random forests, and guided random forests. Moreover, five boosting approaches are considered, namely: Discrete AdaBoost, Real AdaBoost, Gentle AdaBoost, Gradient Boosting and Extreme Gradient Boosting. Gentle AdaBoost is the best performing model on most metrics, except for sensitivity, where Extreme Gradient Boosting performs best. Furthermore, soft majority voting among the models considered is capable of improving the Cohen’s Kappa and the F1 score but does not provide improvements on other metrics.