Globalized Multiple Balanced Subsets With Collaborative Learning for Imbalanced Data.

Area of Science:

• Machine Learning
• Data Science
• Computer Science

Background:

• Imbalanced datasets pose challenges for classifying minority and majority samples.
• Existing balanced bagging methods lack a unified framework and global data perspective.

Purpose of the Study:

• To propose a novel learning framework, globalized multiple balanced subsets with collaborative learning (GMBSCL), to address limitations in imbalanced data classification.
• To enhance collaborative learning across data subsets and integrate global distribution information.

Main Methods:

• Developed a unified learning framework integrating multiple balanced subsets.
• Introduced regularization term R_S for collaborative learning by ensuring consistent minority sample outputs across subsets.
• Introduced regularization term R_W to provide global information to classifiers by aligning subset solution vectors with the entire dataset's direction.

Main Results:

• The proposed GMBSCL framework effectively addresses the limitations of traditional balanced bagging.
• Experimental results demonstrate the superior performance and effectiveness of the GMBSCL approach in imbalanced classification tasks.

Conclusions:

• GMBSCL offers an effective solution for imbalanced data classification by fostering collaborative learning and utilizing global data insights.
• The novel regularization terms R_S and R_W are key to the framework's success in improving classification accuracy.