Feature selection based on multimodal multi-objective particle swarm optimization and prior information

Due to the conflicting objectives of classification accuracy and selected features size, feature selection is typically approached as a multi-objective optimization problem. However, traditional methods often overlook the inherent multimodal nature of feature selection. Additionally, these methods m...

Ausführliche Beschreibung

Gespeichert in:
Bibliographische Detailangaben
Veröffentlicht in:Pattern analysis and applications : PAA Jg. 28; H. 4; S. 181
Hauptverfasser: Liu, Wenkai, Ling, Qinghua, Han, Fei, Han, Henry, Shi, Jinlong
Format: Journal Article
Sprache:Englisch
Veröffentlicht: London Springer London 01.12.2025
Springer Nature B.V
Schlagworte:
Accuracy
Algorithms
Archives & records
Artificial intelligence
Classification
Computer Science
Datasets
Equivalence
Feature selection
Genetic algorithms
Machine learning
Methods
Multiple objective analysis
Mutation
Optimization
Original Article
Particle swarm optimization
Pattern Recognition
Feature selection
Multimodal multiobjective optimization
Niching technique
Particle swarm optimization
ReliefF
ISSN:1433-7541, 1433-755X
Online-Zugang:Volltext
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
Beschreibung
Zusammenfassung:Due to the conflicting objectives of classification accuracy and selected features size, feature selection is typically approached as a multi-objective optimization problem. However, traditional methods often overlook the inherent multimodal nature of feature selection. Additionally, these methods might ignore the importance of filter-based prior knowledge in forming equivalent feature subsets, weakening the ability to search for such subsets. An improved feature selection algorithm, named NRMOPSO, is proposed in this study, which is based on multimodal multi-objective particle swarm optimization and integrates a niche method with ReliefF. Initially, the Incrementally Expanding Niche Strategy (IENS) adjusts niche size for comprehensive initial exploration. Subsequently, the ReliefF algorithm evaluates feature importance, incorporating ReliefF-based prior information into the particle search to include significant unselected features while retaining essential ones. Experimental results on 14 UCI datasets indicate that the proposed algorithm effectively identifies multiple equivalent feature subsets and, on high-dimensional datasets, achieves smaller feature subsets without compromising classification accuracy when compared with five classical and advanced multimodal multi-objective optimization algorithms.
Bibliographie:ObjectType-Article-1
SourceType-Scholarly Journals-1
ObjectType-Feature-2
content type line 14
ISSN:1433-7541
1433-755X
DOI:10.1007/s10044-025-01556-0