Optimal selection of resampling methods for imbalanced data with high complexity.

Area of Science:

• Machine Learning
• Data Science
• Computer Science

Background:

• Class imbalance is a significant challenge in classification tasks, often leading to biased decision boundaries favoring the majority class.
• Data-level solutions like resampling aim to address class imbalance, but can sometimes degrade classification performance due to the overgeneralization problem.

Purpose of the Study:

• To investigate the overgeneralization problem in resampling techniques for imbalanced datasets, particularly in complex data settings.
• To propose and evaluate methods for mitigating overgeneralization and provide guidance on selecting optimal resampling strategies for imbalanced and complex datasets.

Main Methods:

• The study employed simulation studies and real-world data analyses to compare various resampling methods.
• Two primary approaches were investigated: incorporating a filtering method into oversampling and applying undersampling.

Main Results:

• For noncomplex datasets, undersampling emerged as the optimal strategy.
• In complex datasets, applying a filtering method to remove misallocated examples during oversampling proved to be the most effective approach.

Conclusions:

• The optimal resampling method for imbalanced datasets depends on data complexity.
• This research offers valuable insights for researchers to select appropriate resampling techniques, especially for complex and imbalanced data scenarios, thereby enhancing classification accuracy.