Contribution
Nonlinearities in generalized models based on different soundscape datasets
* Presenting author
Abstract:
Research revealed that nonlinear models of the Pleasantness and Eventfulness of soundscapes can have small advantages over linear regression. Consequently, knowledge of the magnitude of potential nonlinear relationships would be beneficial for the choice of an appropriate analysis methods. However, soundscape studies vary in their study design and apply different performance metrices, predictors, and analysis methods, making comparison difficult and challenges the question regarding the necessity and the choice of a specific nonlinear method. Therefore, we have applied different methods and identical performance measures on multiple datasets, aiming at the creation of generalizing models for the use with imbalanced clustered data of small sample sizes which we often face when dealing with soundscape data. Datasets used cover field-based outdoor soundscapes, field-based indoor, lab-based indoor, and lab-evaluated outdoor soundscapes, which differ in sample size and study setting, (e.g., person or location centering). Results revealed performance advantages in out-of-sample R2 from 0.00 to 0.13, and thus suggest that the observed differences in the advantage of nonlinear methods depend strongly on the variance in the data and the amount of group leakage induced in the cross-validation. Herewith, we lay the foundation for uncovering the underlying relationships in more tailored future studies.