A New Representation in PSO for Discretization-Based Feature Selection.

Area of Science:

• Machine Learning
• Data Preprocessing
• High-Dimensional Data Analysis

Background:

• Discretization and feature selection (FS) are crucial for machine learning on high-dimensional data.
• Univariate discretization before FS can discard feature interaction information, potentially degrading performance.
• Previous work showed combining discretization and FS using evolve particle swarm optimization (EPSO) outperformed separate stages.

Purpose of the Study:

• To introduce Potential Particle Swarm Optimization (PPSO), a novel method for integrated discretization and feature selection.
• To enhance the search process through a reduced search space representation and improved fitness function.
• To evaluate PPSO's effectiveness on high-dimensional datasets.

Main Methods:

• Development of the Potential Particle Swarm Optimization (PPSO) algorithm.
• Implementation of a new search space representation within PPSO.
• Design of a novel fitness function for guiding PPSO's search.
• Comparative analysis against two-stage (discretization then FS) and other integrated methods (EPSO, traditional methods).

Main Results:

• PPSO selected less than 5% of features across ten high-dimensional datasets.
• PPSO achieved significantly higher accuracy than the two-stage approach on seven datasets.
• PPSO outperformed or matched EPSO on most datasets, with fewer selected features.
• PPSO demonstrated superior or comparable performance to traditional methods in generalization and learning capacity.

Conclusions:

• PPSO offers an effective integrated approach for discretization and feature selection in machine learning.
• The method significantly reduces feature dimensionality while enhancing classification accuracy.
• PPSO presents a promising alternative to traditional multi-stage preprocessing techniques for high-dimensional data.