A Boundary-Information-Based Oversampling Approach to Improve Learning Performance for Imbalanced Datasets.

Area of Science:

• Machine Learning
• Data Science
• Artificial Intelligence

Background:

• Oversampling is a key data preprocessing technique for imbalanced datasets.
• Traditional classifiers struggle with imbalanced data, necessitating effective oversampling methods.

Purpose of the Study:

• To review existing oversampling techniques, categorizing them as danger-information-based (DIBO) or safe-information-based (SIBO).
• To propose a novel boundary-information-based oversampler (BIBO) that addresses limitations of DIBOs and SIBOs.
• To evaluate the performance of BIBO against DIBO and SIBO methods.

Main Methods:

• Categorization of oversampling techniques into DIBO and SIBO based on sample generation areas.
• Introduction of a new concept of 'boundary information' for sample generation.
• Experimental comparison of DIBO, SIBO, and the proposed BIBO using key classification metrics.

Main Results:

• DIBOs and BIBO showed superior recall and negative class precision compared to SIBOs.
• SIBOs and BIBO demonstrated better specificity and positive class precision than DIBOs.
• BIBO achieved superior performance across integrated metrics compared to both DIBOs and SIBOs.

Conclusions:

• DIBOs risk misclassifying negative examples, while SIBOs can misclassify borderline positive examples.
• The proposed BIBO effectively balances the creation of samples near decision boundaries, mitigating the drawbacks of previous methods.
• BIBO offers an improved approach to handling imbalanced data through strategic oversampling.