Hashing-Based Undersampling Ensemble for Imbalanced Pattern Classification Problems.

Area of Science:

• Machine Learning
• Data Science
• Computer Science

Background:

• Imbalanced classification is a common challenge in machine learning.
• Traditional undersampling methods risk losing valuable majority class data.
• This can negatively impact classifier performance on minority classes.

Purpose of the Study:

• To introduce a novel undersampling technique to mitigate information loss.
• To improve classification performance on imbalanced datasets.
• To present the hashing-based undersampling ensemble (HUE) method.

Main Methods:

• The HUE method utilizes hashing to divide majority class samples into multiple subspaces.
• Each training subset comprises samples from a subspace and adjacent ones.
• An ensemble of classifiers is trained on these diverse subsets with all minority samples.

Main Results:

• The HUE method was evaluated on 25 UCI datasets.
• Performance was compared against existing state-of-the-art undersampling techniques.
• Experimental results demonstrated superior performance of HUE.

Conclusions:

• The hashing-based undersampling ensemble (HUE) effectively handles imbalanced classification problems.
• HUE outperforms other methods, particularly on highly imbalanced datasets.
• The subspace construction prevents the loss of informative majority samples.