A Variable Selection Analysis for Soundscape Emotion Modeling Using Decision Tree Regression and Modern Information Criteria

Abstract

During the last decade, soundscape research has become one of the most active topics in Acoustics. This work provides a nonlinear variable selection analysis over the well-known dataset ‘emo-soundscapes’. Namely, we provide a selection of the soundscape indicators using a nonlinear and nonparametric model as a regression tree method. Modern techniques (proposed in the literature) have been used, first for ranking the variables and then for choosing the effective number of features. We have also compared and discussed our results with those provided previously in the literature. This study, based on modern techniques in selecting the effective number of variables, confirms the result presented in previous recent work (but based on a linear model) that very parsimonious models should be considered (in the case of a nonlinear model, it is based on very few variables, from 2 to 4, depending on the output). All the results are obtained by analyzing a single dataset. As future research works, we plan to extend our study by also considering alternative datasets.

Keywords

• Soundscape emotion recognition
• decision tree regression
• ranking methods
• variable selection