Comparative analysis of instance selection algorithms for instance-based classifiers in the context of medical...

Area of Science:

• Machine Learning
• Computer Science
• Data Science

Background:

• Effective use of available data is crucial for developing robust pattern classifiers.
• Instance-based classifiers rely on a representative set of data points (instances) for optimal performance.

Purpose of the Study:

• To comparatively analyze instance selection algorithms for constructing effective instance-based classifiers.
• To evaluate algorithms based on classification performance, time efficiency, and storage requirements.
• To assess the impact of dataset size on the performance of instance selection methods.

Main Methods:

• Evaluated seven established instance selection algorithms against random selection.
• Utilized a k-nearest neighbor classifier for performance assessment.
• Tested algorithms on simulated Gaussian data and two clinical breast cancer datasets.

Main Results:

• Dataset size was reduced to under 3% of original size with maintained or improved classification performance.
• Random mutation hill climbing demonstrated superior performance among the evaluated algorithms.
• Some existing algorithms performed worse than random selection, highlighting the importance of algorithm choice.

Conclusions:

• Instance selection is beneficial for instance-based classifiers, enhancing performance and reducing resource demands.
• The effectiveness of selection algorithms generally increases with a larger pool of available instances.
• Careful selection of the appropriate instance selection algorithm is critical for achieving optimal results.