A Noise-Filtered Under-Sampling Scheme for Imbalanced Classification.

Area of Science:

• Machine Learning
• Data Science
• Artificial Intelligence

Background:

• Class imbalance is a common problem in machine learning, potentially biasing models toward majority classes.
• Traditional under-sampling methods use all minority data, risking performance degradation due to noisy examples.
• Effective handling of imbalanced data is crucial for reliable classifier performance.

Purpose of the Study:

• To propose a new under-sampling scheme that incorporates a noise filter prior to resampling.
• To enhance the performance of existing under-sampling techniques by mitigating the impact of noisy minority data.
• To evaluate the effectiveness of the proposed scheme across various under-sampling algorithms and imbalanced ratios.

Main Methods:

• A novel under-sampling scheme integrating a noise filtering pre-processing step.
• Implementation and evaluation of the scheme with four popular under-sampling methods: Undersampling + Adaboost, RUSBoost, UnderBagging, and EasyEnsemble.
• Benchmarking and statistical significance analysis to validate performance improvements.

Main Results:

• The proposed noise-filtering under-sampling scheme significantly improves the performance of baseline methods.
• Improvements were observed across key metrics for imbalanced classification: area under the curve (AUC), -measure, and -mean.
• The study also provides insights into the relationship between algorithm performance and the degree of class imbalance.

Conclusions:

• Incorporating a noise filter before under-sampling is an effective strategy for improving classifier performance on imbalanced datasets.
• The enhanced under-sampling methods demonstrate superior results compared to their original counterparts.
• The findings offer a valuable contribution to data preprocessing techniques for handling class imbalance in machine learning.