Selecting critical patterns based on local geometrical and statistical information.

Area of Science:

• Machine Learning
• Pattern Recognition
• Data Science

Background:

• Traditional pattern selection methods are classifier-dependent.
• Effective pattern selection is crucial for training classifiers requiring spatial data.
• Existing methods often lack generalizability across different classifier types.

Purpose of the Study:

• To present a classifier-agnostic pattern selection method.
• To identify critical patterns (edge and border) essential for classifier training.
• To improve the efficiency and accuracy of machine learning models.

Main Methods:

• Pattern selection based on input space location.
• Identification of edge patterns using approximated tangent hyperplanes.
• Identification of border patterns using local probability.
• Evaluation across multiple classifiers: multilayer perceptrons, radial basis functions, support vector machines, and nearest neighbors.

Main Results:

• The method selects patterns representing class boundaries and preserving decision surfaces.
• Achieved consistent accuracy comparable to state-of-the-art approaches.
• Demonstrated effectiveness with reduced datasets.
• Showcased applicability across diverse popular classifiers.

Conclusions:

• The proposed method offers a versatile and effective approach to pattern selection.
• It reduces data requirements while maintaining high and consistent accuracy.
• This technique enhances the training of classifiers that utilize spatial information.