Document Type : Research Manuscript

Authors

• Sahar Abbasi ¹
• Radmin Sadeghian ²
• Maryam Hamedi ²

¹ PhD student, Payam Noor University, PO Box 3697-19395, Tehran, Iran;

² Payame Noor University

Abstract

Multi-label classification assigns multiple labels to each instance, crucial for tasks like cancer detection in images and text categorization. However, machine learning methods often struggle with the complexity of real-life datasets. To improve efficiency, researchers have developed feature selection methods to identify the most relevant features. Traditional methods, requiring all features upfront, fail in dynamic environments like media platforms with continuous data streams. To address this, novel online methods have been created, yet they often neglect optimizing conflicting objectives. This study introduces an objective search approach using mutual information, feature interaction, and the NSGA-II algorithm to select relevant features from streaming data. The strategy aims to minimize feature overlap, maximize relevance to labels, and optimize online feature interaction analysis. By applying a modified NSGA-II algorithm, a set of non-dominant
solutions is identified. Experiments on eleven datasets show that the proposed approach outperforms advanced online feature selection techniques in predictive accuracy, statistical analysis, and stability assessment.

Keywords

• Streaming Data Feature selection
• Online Multi-Label Learning
• Feature interaction
• Multi-Objective Optimization
• Mutual Information

Main Subjects

• Machine Learning