Efficient Selection of Gaussian Kernel SVM Parameters for Imbalanced Data.

Area of Science:

• Bioinformatics
• Machine Learning
• Computational Biology

Background:

• Medical data mining frequently uses class prediction models for data classification.
• High-dimensional gene expression datasets present challenges due to skewed class distributions, leading to poor minority class prediction.
• Support Vector Machines (SVMs) are powerful classifiers but require careful parameter optimization, especially with imbalanced data.

Purpose of the Study:

• To address class imbalance and parameter selection in SVM classifiers for binary classification problems.
• To improve the generalization ability and classification performance of SVMs on imbalanced medical datasets.
• To develop novel methods for adjusting SVM cutoff thresholds and optimizing SVM parameters efficiently.

Main Methods:

• Proposed a novel adjustment method, b-SVM, for modifying the SVM cutoff threshold.
• Introduced Min-max gamma selection, a fast approach for optimizing SVM model parameters without extensive k-fold cross-validation.
• Evaluated algorithms using simulated and real imbalanced medical datasets, comparing against standard SVM and existing methods.

Main Results:

• The proposed b-SVM and Min-max gamma selection algorithms demonstrated superior performance compared to standard SVM and over-sampling techniques.
• Min-max gamma selection was found to be at least 10 times faster than cross-validation selection based on average running times across six real datasets.
• The developed methods effectively improved classification accuracy for minority classes in imbalanced datasets.

Conclusions:

• The novel b-SVM and Min-max gamma selection methods offer significant improvements for SVM classification on imbalanced medical data.
• Min-max gamma selection provides a computationally efficient alternative for SVM parameter optimization.
• These algorithms hold promise for enhancing the identification of marker genes and improving cancer cell classification.