Compressed kNN: K-Nearest Neighbors with Data Compression.

Area of Science:

• Computer Science
• Machine Learning
• Data Mining

Background:

• The k-nearest neighbors (kNN) classification algorithm is widely used but suffers from high memory consumption with large datasets.
• Existing kNN variations, like condensed kNN, aim to reduce dataset size but may not be optimal for all data types.
• Handling categorical data with traditional kNN presents challenges due to its inherent structure and memory demands.

Purpose of the Study:

• To propose a novel, structure-less kNN algorithm variation optimized for categorical data.
• To address the memory consumption limitations of standard kNN algorithms when applied to large categorical datasets.
• To enhance the efficiency of kNN classification for categorical data through data compression.

Main Methods:

• A new kNN algorithm variation designed for structure-less processing of categorical data.
• Implementation of a data compression phase prior to kNN classification.
• On-the-fly decompression of compressed data during the classification process.

Main Results:

• Demonstrated significant reduction in memory requirements for processing large categorical datasets.
• Maintained classification accuracy comparable to traditional kNN methods.
• Observed a slight decrease in overall processing time due to efficient data handling.

Conclusions:

• The proposed kNN variation effectively reduces memory footprint for categorical data classification.
• Data compression is a viable strategy to enhance kNN performance with large categorical datasets.
• This approach offers a practical solution for applying kNN to memory-intensive categorical data problems.