Handling Imbalanced Classification Problems With Support Vector Machines via Evolutionary Bilevel Optimization.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Science

Background:

• Support Vector Machines (SVMs) are widely used for binary classification.
• Traditional SVMs assume equal misclassification costs, which is often unrealistic for imbalanced datasets.
• Real-world data frequently exhibits uneven class distributions, posing challenges for standard SVMs.

Purpose of the Study:

• To introduce EBCS-SVM (evolutionary bilevel cost-sensitive SVMs) for handling imbalanced classification.
• To simultaneously learn support vectors and optimize SVM hyperparameters, including kernel parameters and misclassification costs.
• To address the limitations of traditional SVMs in scenarios with uneven class distributions.

Main Methods:

• EBCS-SVM employs a bilevel optimization approach.
• The upper level uses an evolutionary algorithm (EA) to optimize hyperparameters.
• The lower level utilizes sequential minimal optimization (SMO) to determine support vectors, with nested optimization.
• The lower level is initialized using successful solutions from previous iterations.

Main Results:

• EBCS-SVM was evaluated on 70 imbalanced classification datasets.
• Performance was compared against several state-of-the-art methods.
• Bayesian testing supported the experimental findings.

Conclusions:

• EBCS-SVM demonstrates significant effectiveness in handling highly imbalanced datasets.
• The proposed method offers an improved approach to SVM hyperparameter and support vector optimization for skewed data.
• The nested evolutionary and SMO optimization strategy proves beneficial for cost-sensitive classification.