Online Nonlinear AUC Maximization for Imbalanced Data Sets

Area of Science:

• Machine Learning
• Data Mining
• Artificial Intelligence

Background:

• Classifying binary imbalanced streaming data presents significant challenges.
• Existing online area under the ROC curve (AUC) maximization methods yield linear classifiers, struggling with data nonlinearity and heterogeneity.

Purpose of the Study:

• To propose a novel algorithm, kernelized online imbalanced learning (KOIL), for nonlinear classification of imbalanced streaming data.
• To enhance the AUC maximization approach by incorporating kernelization for improved performance on complex datasets.

Main Methods:

• Developed the KOIL algorithm, which maximizes AUC score while minimizing a functional regularizer to produce nonlinear classifiers.
• Introduced two fixed-budget buffers to manage support vectors and capture global decision boundary information.
• Implemented a strategy to confine the influence of new support vectors to their k-nearest opposite support vectors for smooth updating.
• Proposed a compensation scheme to prevent information loss when buffers are full, ensuring performance comparable to infinite budgets.
• Utilized multiple kernel learning to automatically learn optimal kernels for data similarity representation.

Main Results:

• The KOIL algorithm demonstrates efficacy in classifying nonlinear and heterogeneous imbalanced streaming data.
• Experimental results on synthetic and real-world datasets validate the proposed approach's performance.
• The buffer management and compensation schemes effectively control support vector count without performance degradation.

Conclusions:

• The KOIL algorithm offers a robust solution for imbalanced streaming data classification, outperforming linear methods.
• The integration of kernelization, buffer management, and multiple kernel learning significantly advances online imbalanced learning capabilities.