Feature selection for machine-learned crowd reactions at collegiate basketball games

Crowds at basketball games react acoustically to events on the court in many ways, including applauding, chanting, cheering, and making distracting noises. Acoustic features can be extracted from recordings of crowd noise at Brigham Young University basketball games to train machine learning models for classifying crowd reactions. In this study, feature selection using random forests and LASSO logistic regression was used to identify the most useful acoustic features for identifying and classifying crowd reactions. Features related to specific 1/3-octave band shapes, sound level, and tonality are found to be highly relevant. Including feature histories can increase classifier accuracies by up to 12%. Interestingly, some features are better predictors of future crowd reactions than current reactions. Reduced feature sets are human-interpretable on a case-by-case basis for the crowd reactions they predict.