Explainable machine learning models for classifying reactions within crowd noise during men's collegiate basketball...

Area of Science:

• Acoustics
• Machine Learning
• Sports Analytics

Background:

• Crowd noise at collegiate basketball games is diverse, including applause, chants, and cheers.
• Acoustic analysis of crowd sounds can inform various applications, from player contracts to fan experience and safety.
• Understanding crowd acoustics is crucial for developing effective crowd monitoring and management systems.

Purpose of the Study:

• To identify key acoustic features for classifying crowd reactions in basketball games using machine learning.
• To evaluate the impact of feature selection on model performance and interpretability.
• To assess the value of incorporating short-term feature temporal histories into classification models.

Main Methods:

• Acoustic features were extracted from crowd recordings during basketball games.
• Machine learning models, including random forests and logistic regression, were employed for classification.
• Feature selection techniques were applied to identify the most predictive acoustic features.

Main Results:

• Features related to 1/3-octave band shapes, sound level, and tonality were found to be highly relevant.
• Incorporating short-term feature temporal histories improved classifier accuracies by up to 12%.
• Some acoustic features proved to be better predictors of future crowd reactions than current ones.

Conclusions:

• Feature selection enhances the interpretability and performance of crowd reaction classification models.
• Temporal dynamics of acoustic features are significant for accurately classifying crowd behavior.
• The study provides insights into the acoustic characteristics of crowd noise for practical applications.