Feature reduction of crowd noise used for machine learning classification

Abstract

This paper discusses methods used to identify the relative importance of audio features in a supervised machine learning (ML) model for predicting crowd behavior at collegiate basketball games. This work builds upon previous research done at Brigham Young University in which ML classifiers were trained using audio recordings of crowds at collegiate basketball games. Previous classifiers were built using several hundred features. A future goal is real-time classification of crowd behavior. This requires reduced computational time for calculating features and classifying audio data. Feature reduction should decrease the computational time for both of these tasks. The audio features can be separated into two categories: (1) spectral features and (2) low-level signal parameters. This paper discusses feature reduction methods—such as random forest Gini importance and p-value feature selection—and compares reduced feature sets. The number of features used in crowd noise classification can be reduced to 10-15 features before seeing a significant decrease in prediction accuracy.