Embedded Feature Selection for Multi-label Classification of Music Emotions

Guo-Zheng Li,Jiaming Liu,Mingyu You,Yan Chen

doi:10.1080/18756891.2012.718113

Guo-Zheng Li, Jiaming Liu + Show 2 more

Open Access

https://doi.org/10.1080/18756891.2012.718113

Copy DOI

Abstract

Abstract When detecting of emotions from music, many features are extracted from the original music data. However, there are redundant or irrelevant features, which will reduce the performance of classification models. Considering the feature problems, we propose an embedded feature selection method, called Multi-label Embedded Feature Selection (MEFS), to improve classification performance by selecting features. MEFS embeds classifier and considers the label correlation. Other three representative multi-label feature selection methods, known as LP-Chi, max and avg, together with four multi-label classification algorithms, is included for performance comparison. Experimental results show that the performance of our MEFS algorithm is superior to those filter methods in the music emotion dataset.

Highlights

In daily life, music plays an important role
As we experience in our daily life, more than one emotion may be evoked by music simultaneously
To select the classifier specified features without the high time cost like wrapper methods, we propose a tradeoff method by introducing an embedded feature selection method into multi-label classification

Summary

Introduction

It influences people emotional by nature, makes people feel happy or sad, angry or relaxed. The problem of automatically categorizing music into emotions was modeled as single-label classification[1,2] or regression[3]. As we experience in our daily life, more than one emotion may be evoked by music simultaneously. In this case, classification and regression with singlelabel can hardly model the multiplicity in music emotion studies. Multi-label approaches are more appropriate in modeling music emotions[4,5]

Methods

Results

Conclusion