A Novel Multiobjective Genetic Programming Approach to High-Dimensional Data Classification.

Area of Science:

• Computer Science
• Machine Learning
• Bioinformatics

Background:

• High-dimensional data from sensing technologies challenge machine learning models.
• Genetic programming (GP) faces issues with solution diversity, multiclass imbalance, and large feature spaces in high-dimensional data.

Purpose of the Study:

• To develop a problem-specific multiobjective GP framework (PS-MOGP) for high-dimensional data classification.
• To address the limitations of existing GP methods in handling large feature spaces and imbalanced multiclass data.

Main Methods:

• Incorporated recursive feature elimination using evolved GP solutions to manage large solution spaces.
• Proposed a progressive domination Pareto archive evolution strategy (PD-PAES) for solution diversity.
• Developed BD with a similar positive and negative class size (BD-SPNCS) to mitigate multiclass imbalance.

Main Results:

• The PS-MOGP framework effectively reduces the solution space in high-dimensional datasets.
• PD-PAES maintained superior solution diversity compared to standard methods.
• BD-SPNCS improved the handling of imbalanced classes in multiclass problems.

Conclusions:

• The PS-MOGP framework demonstrates superior performance in high-dimensional data classification.
• The proposed methods offer effective solutions for diversity, feature space reduction, and class imbalance in GP.
• PS-MOGP outperforms state-of-the-art methods on benchmark and real-world datasets.