Comparative Analysis of Machine Learning Algorithms For Classifying Music According to Mood

Date of Award

7-1-2023

Document Type

Thesis

Degree Name

Master of Science in Computer Science

First Advisor

Andrei D. Coronel, PhD

Abstract

This research compares feature set variations and several methods of machine learning techniques for the classification of music according to mood. The study tests a dataset which was created by gathering annotations of mood from the AllMusic website. This entails the creation of feature sets using three different ways: a feature set based on music principles, a feature set built using Kendall’s Tau correlation on music extracted by the Symbolic software, and feature set built by using a Decision Tree as a feature selection tool prior to classification. These feature sets were then tested using SVM, Naive Bayes and decision tree algorithms, using accuracy and F-measure as metrics of performance. This study also repeats the same methodology for the classification of music tracks that may be labeled under two mood categories to test the performance of the models created. Low results were seen from the first two feature sets used in the study so experiments were performed in an attempt to raise these results. Promising results were seen in the third feature set which was built using decision tree as a feature selection tool. The best feature set in these experiments resulted with a classification accuracy of 82.0% in the initial splitting for data points and an F-measure of 81.3%, using the Decision Tree when classifying songs with low valence in the context of the dual mood label classification.

Link to Full Text

Share

COinS