Granular Ball Sampling for Noisy Label Classification or Imbalanced Classification.

Area of Science:

• Machine Learning
• Data Mining
• Artificial Intelligence

Background:

• Classification tasks often face challenges with large datasets and noisy labels.
• Existing data sampling methods may not effectively handle data reduction while preserving or improving classification accuracy.

Purpose of the Study:

• To introduce a general sampling method, granular-ball sampling (GBS), for classification problems.
• To demonstrate GBS's ability to reduce data size and improve data quality, particularly in noisy label classification.
• To evaluate GBS's effectiveness as an undersampling method for imbalanced classification.

Main Methods:

• The granular-ball sampling (GBS) method utilizes adaptively generated hyperballs to cover the data space.
• Data points on these hyperballs are selected for the sampled dataset.
• GBS is designed to accurately describe data boundaries.

Main Results:

• GBS reduces data size while improving data quality in noisy label classification.
• GBS achieves classification accuracy comparable to original datasets and significantly higher than random sampling.
• GBS demonstrates effectiveness as an undersampling technique for imbalanced classification.
• The method exhibits a time complexity close to O(N), accelerating classifier performance.

Conclusions:

• Granular-ball sampling (GBS) is a versatile and effective data reduction technique for classification.
• GBS offers significant advantages over random sampling, particularly for noisy and imbalanced datasets.
• The open-source GBS library facilitates its application in diverse classification tasks.