Diversified Sensitivity-Based Undersampling for Imbalance Classification Problems.

Area of Science:

• Computer Science
• Machine Learning
• Data Science

Background:

• Imbalanced pattern classification poses challenges for machine learning models.
• Existing undersampling methods like random resampling or boundary-based resampling have limitations.
• These methods often ignore data distribution and informative samples, impacting model performance.

Purpose of the Study:

• To propose a novel diversified sensitivity-based undersampling method.
• To address limitations of current undersampling techniques in handling imbalanced datasets.
• To improve classifier sensitivity and generalization capability in imbalanced classification tasks.

Main Methods:

• Clustering majority class samples to capture distribution information and enhance diversity.
• Applying a stochastic sensitivity measure for selecting informative samples from both majority and minority classes.
• Iterative clustering and sampling to achieve a balanced dataset with high classifier sensitivity.

Main Results:

• The proposed method effectively captures data distribution and enhances sample diversity.
• It selects informative samples from both majority and minority classes, improving balance.
• Evaluated on 14 UCI datasets, the method demonstrated good generalization capability.

Conclusions:

• The diversified sensitivity-based undersampling method offers an effective approach for imbalanced pattern classification.
• It outperforms traditional methods by considering data distribution and sample informativeness.
• The method shows promise for improving the performance of classifiers on imbalanced datasets.