Selective oversampling approach for strongly imbalanced data.

Area of Science:

• Machine Learning
• Data Science
• Computer Science

Background:

• Imbalanced datasets present significant challenges in various real-world machine learning applications.
• Oversampling is a common technique to address class imbalance by increasing the number of minority class instances.

Purpose of the Study:

• To introduce a novel Selective Oversampling Approach (SOA) for improving classifier performance on imbalanced data.
• To enhance existing oversampling methods by focusing on representative minority samples.

Main Methods:

• The proposed Selective Oversampling Approach (SOA) identifies representative minority class samples using outlier detection.
• These selected samples are then used for synthetic oversampling to augment the minority class.
• The SOA was integrated with two state-of-the-art methods: Synthetic Minority Oversampling Technique (SMOTE) and Adaptive Synthetic Sampling (ADASYN).

Main Results:

• The Selective Oversampling Approach (SOA) demonstrated improved performance when combined with SMOTE and ADASYN.
• Evaluations on four synthetic and four real-world datasets consistently showed that SOA-enhanced methods achieved equal or superior predictive performance.
• The outlier detection component effectively isolates crucial minority samples for more effective synthetic data generation.

Conclusions:

• The Selective Oversampling Approach (SOA) is an effective strategy for enhancing oversampling techniques in imbalanced learning scenarios.
• SOA offers a robust method to improve classifier accuracy by focusing synthetic data generation on the most informative minority instances.
• The proposed method provides a valuable advancement for handling imbalanced data in practical machine learning applications.