Bayesian network-based over-sampling method (BOSME) with application to indirect cost-sensitive learning.

Area of Science:

• Machine Learning
• Data Science
• Artificial Intelligence

Background:

• Supervised learning algorithms struggle with imbalanced datasets, often misclassifying minority classes.
• Existing over-sampling methods generate synthetic data but may not capture underlying class distributions effectively.

Purpose of the Study:

• To introduce a novel over-sampling technique, Bayesian network-based over-sampling method (BOSME), for imbalanced data classification.
• To evaluate BOSME's performance against the Synthetic Minority Over-sampling Technique (SMOTE) in cost-sensitive learning scenarios.

Main Methods:

• Learned Bayesian networks from minority class instances to model their probability distribution.
• Generated synthetic minority class instances guided by the learned Bayesian network's probability distribution.
• Compared BOSME with SMOTE using state-of-the-art classifiers and various datasets under indirect cost-sensitive learning.

Main Results:

• BOSME demonstrated statistically significant improvements over SMOTE in terms of expected misclassification cost.
• The proposed method showed enhanced ability to predict the minority class in imbalanced datasets.
• Experiments confirmed BOSME's effectiveness across different classifiers and datasets.

Conclusions:

• BOSME offers a superior approach to handling imbalanced data compared to traditional methods like SMOTE.
• The Bayesian network-based instance generation effectively preserves and utilizes the minority class distribution for better classification.
• This methodology provides a promising direction for improving machine learning model performance on imbalanced datasets.