Robust Distance Measures for kNN Classification of Cancer Data.

Area of Science:

• Machine Learning
• Bioinformatics
• Computational Biology

Background:

• The k-Nearest Neighbor (kNN) classifier is a versatile machine learning algorithm.
• Its performance relies heavily on the chosen distance or similarity measure.
• Evaluating distance measures is crucial for optimizing kNN in biomedical data analysis.

Purpose of the Study:

• To investigate the impact of various distance measures on kNN classifier performance.
• To compare common and novel distance metrics, including Sobolev and Fisher, on cancer datasets.
• To establish a robust ranking of distance measures for kNN in biomedical classification.

Main Methods:

• Applied kNN classification to four distinct cancer datasets.
• Utilized a range of common and novel distance measures, such as Hassanat, Manhattan, Sobolev, and Fisher.
• Evaluated and ranked the performance of each distance measure across the datasets.

Main Results:

• Novel distance measures, particularly Sobolev, demonstrated performance comparable to established metrics.
• Hassanat, Sobolev, and Manhattan measures exhibited robust performance across multiple datasets.
• Some distance measures showed dataset-specific efficacy, highlighting sensitivity to data characteristics.

Conclusions:

• The choice of distance measure significantly influences kNN classifier performance in cancer data analysis.
• Novel measures like Sobolev offer competitive alternatives to traditional metrics.
• Benchmarking distance measures on relevant data is essential for selecting the optimal metric for specific classification tasks.