A Training Data Set Cleaning Method by Classification Ability Ranking for the k -Nearest Neighbor Classifier.

Area of Science:

• Computer Science
• Machine Learning
• Pattern Recognition

Background:

• K-nearest neighbor (KNN) is a popular supervised classification technique.
• KNN performance is sensitive to noisy or incorrect samples in training data.
• Training data cleaning (TDC) is crucial for enhancing KNN accuracy.

Purpose of the Study:

• To propose a novel TDC method based on classification ability ranking (CAR) for KNN classifiers.
• To improve the performance and reduce the computational complexity of KNN classifiers through data cleaning.

Main Methods:

• Developed a classification ability function to rank training samples based on their contribution to KNN classification.
• Utilized a leave-one-out (LV1) strategy to assess sample misclassification potential.
• Implemented a CAR-based TDC method to remove low-classification-ability samples.

Main Results:

• The CAR-based TDC method significantly reduced classification error rates across ten real-world datasets.
• The method effectively identified and removed noisy or erroneous training samples.
• A smaller, cleaned training dataset led to reduced computational complexity.

Conclusions:

• The proposed CAR-based TDC method is effective in enhancing KNN classifier performance.
• This approach offers a robust solution for improving the accuracy and efficiency of KNN-based pattern classification.