An oversampling method for imbalanced data based on spatial distribution of minority samples SD-KMSMOTE.

Area of Science:

• Data Science
• Machine Learning
• Artificial Intelligence

Background:

• Data imbalance is a significant challenge in fields like medicine and finance.
• Existing oversampling methods often lack effectiveness due to random sampling or limited scope.
• These limitations negatively impact classification performance.

Purpose of the Study:

• To propose a novel imbalanced data oversampling method, SD-KMSMOTE.
• To address the limitations of current oversampling techniques by incorporating spatial distribution analysis.
• To improve classification accuracy on imbalanced datasets.

Main Methods:

• Developed SD-KMSMOTE, an oversampling method based on minority sample spatial distribution.
• Incorporated filter noise pre-treatment and near-neighbor category information to remove minority class noise.
• Designed a new sample synthesis method with weight calculation rules considering spatial distribution and clustering of minority samples.

Main Results:

• SD-KMSMOTE demonstrated superior performance compared to existing methods.
• Improvements were observed across key metrics including precision, recall, F1 score, G-mean, and AUC.
• The method effectively expanded imbalanced datasets in medical and other domains.

Conclusions:

• SD-KMSMOTE offers a significant advancement in handling imbalanced data.
• The method's focus on spatial distribution and noise reduction leads to enhanced classification outcomes.
• SD-KMSMOTE provides a robust solution for imbalanced datasets across various applications.