Exploratory undersampling for class-imbalance learning.

Area of Science:

• Machine Learning
• Data Science
• Artificial Intelligence

Background:

• Class imbalance is a common problem in machine learning.
• Undersampling is an efficient but data-losing method for handling class imbalance.
• Existing methods often ignore valuable majority class data.

Purpose of the Study:

• To propose novel algorithms for addressing class imbalance.
• To overcome the data deficiency of traditional undersampling techniques.
• To improve the performance of imbalanced learning models.

Main Methods:

• EasyEnsemble: Samples multiple subsets of the majority class, trains learners on each, and combines their outputs.
• BalanceCascade: Sequentially trains learners, removing correctly classified majority instances at each step.
• Both methods aim to retain more majority class information.

Main Results:

• Both EasyEnsemble and BalanceCascade achieve higher Area Under the ROC Curve (AUC), F-measure, and G-mean.
• Performance surpasses existing class-imbalance learning methods.
• Training time is comparable to undersampling with an equal number of weak classifiers.

Conclusions:

• EasyEnsemble and BalanceCascade are effective solutions for class imbalance problems.
• These algorithms offer improved predictive performance without sacrificing training efficiency.
• They represent a significant advancement over traditional undersampling techniques.