Serial Number :  http://dx.doi.org/10.14419/ijet.v7i4.39.23697 

(https://www.sciencepubco.com/index.php/ijet/article/view/23697)

Title : Toward Accurate Music Classification Using Local Set-based Multi-label prototype Selection

Author : Seo, Wang-duk, Seo, Sang-hyun, Lee, Jae-sung

Vol. : Vol.7 No.4 [2018]

Pages : 52-55

Published by : Science Publishing Corporation 

Date : 2018.10

Register Information : SCOPUS

etc. : First Author=Lee, Jae-sung(Seo, Wang-duk); Corresponding author=Lee, Jae-sung, Co-author= Seo, Sang-hyun

<Abstract>

Background/Objectives: Multiple music tags enable quick searching and selection of music clips for by end-users to listen to. Our goal is to improve the accuracy of automatic music categorization.

Methods/Statistical analysis: We propose a local set-based multi-label prototype selection to remove noisy samples in datasets without transforming multi-label datasets to single-label datasets by searching the local set of each sample. To validate the superiority of the proposed method, we use ten multi-label music datasets and Hamming loss as a performance measurement, which counts the symmetric difference between predicted labels and ground truth labels.

Findings: Considering time and cost, manual categorization of a large collection of music clips is generally impractical. As such, an automated approach for addressing this task through the training of music tags annotated from an online system is employed. In the real world, multiple labels can be annotated to a music clip by users of an online system, resulting in unintended noisy samples due to inaccurate annotations. Conventional methods attempt to transform multi-label datasets to single-label datasets that can yield additional computational cost and unintended removal of non-noisy samples. In this paper, we propose a novel prototype selection method for multi-label music categorization. Experimental results indicate that the proposed method performed the best performance on nine music datasets. From the experiment of CAL500 dataset, multi-label classification performance based on the training samples selected by our proposed method was 0.1402, which indicates that 15,020 labels on average were correctly classified for 100 test samples. Compared to the second-best performance by the compared method, our method was able to classify 245 more labels.

Improvements/Applications: Experiments using ten different musical datasets showed that the proposed method demonstrated better performance than the compared methods.

<Key words>

 Auto Music Tags Annotation, Multi-label Learning, Prototype Selection, Pruned Problem Transformation, Local Set-based Smoother