Particle Swarm Optimized Hybrid Kernel-Based Multiclass Support Vector Machine for Microarray Cancer Data Analysis.

Area of Science:

• Bioinformatics
• Computational Biology
• Machine Learning in Oncology

Background:

• Microarray-based cancer classification presents challenges due to imbalanced datasets and the complexity of optimal decision models.
• Multiclass Support Vector Machines (MCSVM) are valuable but their performance is sensitive to penalty factors, kernel types, and parameters.

Purpose of the Study:

• To propose an improved model, PSO-PCA-LGP-MCSVM, for enhanced performance in imbalanced microarray-based cancer classification.
• To introduce a novel hybrid kernel, Linear-Gaussian-Polynomial (LGP), for MCSVM that combines advantages of linear, Gaussian, and polynomial kernels.

Main Methods:

• The proposed model integrates Particle Swarm Optimization (PSO) and Principal Component Analysis (PCA) with MCSVM.
• A novel hybrid LGP kernel is developed, linearly combining linear and Gaussian kernels while embedding a polynomial kernel, and its validity is proven.
• Comparative experiments were conducted on two dual and two multiclass imbalanced microarray datasets.

Main Results:

• The PSO-PCA-LGP-MCSVM model demonstrated superior performance compared to three single-kernel based models (linear, Gaussian, polynomial).
• Evaluation metrics including F-score, G-mean, and Accuracy confirmed the model's enhanced global feature extraction, prediction, and learning abilities.
• The hybrid LGP kernel proved effective in improving classification accuracy on imbalanced datasets.

Conclusions:

• The PSO-PCA-LGP-MCSVM model offers a significant advancement for imbalanced microarray-based cancer classification.
• The hybrid LGP kernel is a key innovation, effectively leveraging the strengths of multiple kernel types.
• This approach provides a more robust and accurate decision model for complex biological data analysis.