Multi-label learning based on positive label correlations using predictive apriori

Multi-label Learning (MLL) is a general task in data mining that consists of three main tasks: classification, label ranking, and multi-label ranking. MLL is a challengeable task due to the problem of the large search space, which is the result of the existing correlations among the labels. Conseque...

Full description

Saved in:
Bibliographic Details
Main Author: Al Azaidah, Raed Hasan Saleh
Format: Thesis
Language:eng
eng
eng
Published: 2019
Subjects:
Online Access:https://etd.uum.edu.my/9022/1/s901576_01.pdf
https://etd.uum.edu.my/9022/2/s901576_02.pdf
https://etd.uum.edu.my/9022/3/s901576_references.docx
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Multi-label Learning (MLL) is a general task in data mining that consists of three main tasks: classification, label ranking, and multi-label ranking. MLL is a challengeable task due to the problem of the large search space, which is the result of the existing correlations among the labels. Consequently, the predictive performance of the existing multi-label classifiers is still low, in comparative of single label classifiers .Overall, this study investigates different problems related to MLL such as the huge loss of information resulted from the transformation step. The ignorance of the correlations among labels in the transformation step has caused limited exploitation of the captured correlations in discovering the applicability of Associative Classification (AC) in Multi-label Classification(MLC). The main objective of this study is to enhance the predictive performance of the classification and ranking tasks in MLL through capturing positive correlations among labels. A new Problem Transformation Method (PTM) that considers positive correlations among labels has been proposed in this study. Along with that, two algorithms have been constructed by capturing the positive correlations where the first algorithm (MLR-PC) captures the positive global correlations and the second algorithm (MLCBA) proposes an adaption of AC algorithm to handle MLC based on the positive local correlations. Seven different datasets in various sizes with diverse characteristics have been chosen to thoroughly conduct this study. The proposed PTM based on the positive correlations among labels has shown superior performance compared to the existing PTMs that only consider the frequency of labels as a transformation criterion. On multi label datasets with high cardinality, the results outperform in relation to datasets with low cardinality. Capturing the positive correlations among labels helps in reducing the large search space in MLL, and hence, improve the predictive performance of the classification and ranking tasks, especially in datasets with high cardinality. Finally, utilizing AC in MLC when considering an appropriate discretization technique, has a high positive influence on solving the problem of MLC.