Random Space Division Sampling for Label-Noisy Classification or Imbalanced Classification.

Area of Science:

• Machine Learning
• Data Mining
• Pattern Recognition

Background:

• Classification tasks often face challenges with large datasets and noisy labels.
• Existing sampling methods may not be universally applicable or efficient.
• Distinguishing between noisy, inner, and boundary points is crucial for effective sampling.

Purpose of the Study:

• To introduce a novel, general-purpose sampling method for classification tasks.
• To enhance classification accuracy, particularly in label-noisy scenarios.
• To reduce dataset size and accelerate classifier performance.

Main Methods:

• Developed "random space division sampling" (RSDS) based on random space division.
• Implemented RSDS to efficiently distinguish and extract boundary points.
• Evaluated RSDS as a general sampling technique and for imbalanced classification.

Main Results:

• RSDS effectively reduces data size while enhancing classification accuracy.
• The method demonstrates superior performance in label-noisy classification tasks.
• RSDS shows efficiency and generalizability across diverse datasets and classifiers.
• Experimental results validate the effectiveness and efficiency of RSDS.

Conclusions:

• RSDS is a versatile and efficient sampling method for classification.
• It offers significant improvements in accuracy and data reduction, especially for noisy data.
• The method's general applicability and online acceleration capabilities make it a valuable tool.