Improving learning from the complex multi-class imbalanced and overlapped data by mapping into higher dimension using...

Area of Science:

• Machine Learning
• Computer Science
• Data Science

Background:

• Traditional classifiers struggle with multi-class problems due to imbalanced samples and data overlap.
• These issues degrade classifier efficiency, especially with increasing class numbers and overlapping attributes.
• The combined effects of imbalanced data and overlapping samples receive limited research attention.

Purpose of the Study:

• To introduce SVM++, a modified Support Vector Machine (SVM) algorithm, designed to improve multi-class classification.
• To address the challenges posed by imbalanced datasets and overlapping sample attributes in machine learning models.
• To enhance classifier performance in scenarios with unequal sample distribution and complex data structures.

Main Methods:

• The proposed SVM++ algorithm involves a three-step process: splitting data into overlapping and non-overlapping sets.
• Algorithm-2 further categorizes overlapped data into Critical-1 and Critical-2 regions, identifying problematic samples.
• A novel kernel mapping function is employed, enhancing traditional SVM by mapping Critical-1 samples to a higher dimension based on distance metrics.

Main Results:

• SVM++ demonstrated superior performance compared to state-of-the-art classifiers across thirty real-world datasets.
• The algorithm effectively handled datasets with varying degrees of sample imbalance and attribute overlap.
• The proposed method significantly improved classification accuracy in challenging multi-class scenarios.

Conclusions:

• SVM++ offers a robust solution for multi-class classification problems plagued by imbalanced data and sample overlap.
• The enhanced kernel mapping and data partitioning strategy are key to SVM++'s improved performance.
• This research highlights the importance of addressing combined data imbalance and overlap for effective classifier development.