A novel template reduction approach for the K-nearest neighbor method.

Area of Science:

• Computer Science
• Machine Learning
• Pattern Recognition

Background:

• K-nearest neighbor (KNN) is a popular classification algorithm.
• Large datasets pose computational challenges for KNN classification.
• Condensing approaches aim to reduce computational load by removing redundant data.

Purpose of the Study:

• Propose a novel condensing algorithm for KNN.
• Reduce the number of training prototypes while maintaining classification accuracy.
• Develop a computationally efficient method for large-scale pattern classification.

Main Methods:

• Introduced a "chain" concept: a sequence of nearest neighbors from alternating classes.
• Identified patterns near the classification boundary using the chain.
• Implemented a cutoff strategy to retain essential training patterns.

Main Results:

• The proposed algorithm significantly reduces the number of training prototypes.
• Maintained classification accuracy comparable to traditional KNN.
• Demonstrated simplicity and speed compared to existing methods.

Conclusions:

• The new condensing algorithm effectively addresses KNN's computational demands for large datasets.
• Offers a practical and efficient solution for pattern classification.
• Provides a valuable alternative for scenarios requiring fast and accurate KNN implementations.