Post-boosting of classification boundary for imbalanced data using geometric mean.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Science

Background:

• Imbalanced datasets pose significant challenges in binary classification tasks.
• Traditional accuracy metrics can be misleading in imbalanced scenarios.
• Existing imbalance learning methods often struggle to balance performance across classes.

Purpose of the Study:

• To introduce a novel imbalance learning method, Post-Boosting of classification boundary for Imbalanced data (PBI).
• To enhance the performance of neural network (NN) classification boundaries on imbalanced data.
• To provide a statistically sound approach using the geometric mean (G-mean) for imbalanced data evaluation.

Main Methods:

• PBI involves two steps: initial NN learning followed by boundary adjustment using PBI.
• The geometric mean (G-mean) is employed to balance accuracies of minority and majority classes.
• A new metric, Majority loss/Minority advance ratio (MMR), is proposed to evaluate class-specific performance.

Main Results:

• PBI significantly improves minority class accuracy while maintaining or improving majority class accuracy.
• The method demonstrates effectiveness on large datasets with high imbalance ratios (up to 0.001).
• Experimental results show PBI outperforms other imbalance learning methods across diverse datasets.

Conclusions:

• PBI offers a robust and effective solution for binary classification with imbalanced data.
• The proposed G-mean metric and MMR provide better insights into model performance on imbalanced datasets.
• PBI represents a significant advancement in improving neural network classification boundary performance.