Imbalanced data preprocessing techniques for machine learning: a systematic mapping study.

Area of Science:

• Computer Science
• Data Science
• Artificial Intelligence

Background:

• Machine Learning (ML) algorithms are increasingly used in domains with imbalanced data.
• Existing solutions involve data preprocessing, cost-sensitive learning, and ensemble methods to mitigate bias.
• These techniques aim to reduce the inherent bias towards the majority class in ML models.

Purpose of the Study:

• To systematically map and analyze research on sampling techniques for ML in imbalanced data applications.
• To identify common preprocessing techniques and ML models used in the field.
• To evaluate the performance of different approaches and suggest future research directions.

Main Methods:

• A systematic mapping methodology was employed to review 9927 papers from 7 digital libraries.
• A filtering process identified 35 representative papers across health, finance, and engineering domains.
• Quantitative analysis was performed on the selected papers to develop taxonomies.

Main Results:

• Oversampling and classical ML models are the most prevalent techniques and models, respectively.
• Neural networks and ensemble ML models demonstrate the highest performance.
• Hybrid sampling techniques show potential for improved results.
• Simulation-based synthetic oversampling has not been applied in the reviewed literature.

Conclusions:

• Classical ML and oversampling are common but not always optimal for imbalanced data.
• Advanced models like neural networks and ensemble methods, combined with hybrid sampling, offer superior performance.
• Simulation-based synthetic oversampling represents a promising avenue for future research in imbalanced data preprocessing.