•  
  •  
 

ASEAN Journal on Science and Technology for Development

Abstract

A multi-label pattern classification system tries to predict the set of class labels of a test example by learning from the training examples with the relevant label sets. Classification that involve datasets having multiple labels found immense of applications in pattern analysis tasks involving image, music and video. A test sample can be labeled to indicate different objects, people, music categories or concepts. Classification problems involving data having multiple labels, have to consider training dataset associated with variety of labels. Multi-label extensions of popular algorithms, kNN and Support Vector Machine (SVM) called Multi-label kNN (ML-kNN) and Ranking-SVM are commonly used for the classification of datasets having multiple labels. It is experimentally established that the distance metric used in the ML-kNN classifier influences its performance and also the kernels used in Ranking-SVM have significant influence in the performance. A learnt distance metric in the ML-kNN method and a kernel learnt using the distance metric learning in the Ranking-SVM is proposed in this paper. The performance is systematically evaluated on six bench-mark datasets using five performance metrics, namely, Hamming loss between the label-vectors, ranking loss, average precision, coverage, and one-error. It is experimentally established that learned distance metrics in place of default distances give a better performance for classifiers handling datasets having multiple labels.

Keywords

Multi-label classification · Distance metric learning · LMDMML- kNN · DMLK-Ranking-SVM · Logdet divergence · Bregman divergence

Publication Date

2025

Received Date

16/10/2024

Revised Date

06/12/2024

Accepted Date

07/01/2025

Download

Share

COinS