An oversampling method for multi-class imbalanced data based on composite weights.

Area of Science:

• Machine Learning
• Data Science
• Artificial Intelligence

Background:

• Imbalanced multi-class datasets pose significant challenges in machine learning, often leading to biased models and poor classification accuracy for minority classes.
• Traditional oversampling techniques may fail to preserve the underlying data distribution or introduce noise, limiting their effectiveness.

Purpose of the Study:

• To develop an advanced oversampling algorithm for imbalanced multi-class datasets that enhances classification accuracy.
• To create a method that effectively balances data quantity while preserving the distribution and information properties of original samples.

Main Methods:

• A novel oversampling algorithm is proposed, utilizing classification ranking and weight setting.
• Data is sorted based on distance to the hyperplane, with iterative intra-class and inter-class sampling at class boundaries.
• Sampling weights are determined by data density and sorting, followed by information assignment to synthetic data.

Main Results:

• The proposed algorithm successfully balances the quantity of multi-class imbalanced data.
• Newly generated samples retain the distribution and information characteristics of the original data.
• Experimental results on UCI imbalanced datasets show classification accuracy around 90%, outperforming SMOTE and SVMOM.

Conclusions:

• The developed oversampling method demonstrates high practicability and generalizability for imbalanced multi-class samples.
• This approach offers a robust solution for improving classification performance in scenarios with skewed data distributions.