Selecting discrete and continuous features based on neighborhood decision error minimization.

Area of Science:

• Machine Learning
• Pattern Recognition
• Data Science

Background:

• Feature selection is crucial for effective pattern recognition and machine learning.
• Estimating classification complexity is a key challenge in developing feature selection algorithms.
• Existing methods may not adequately handle mixed data types or accurately assess subspace complexity.

Purpose of the Study:

• To propose a novel feature evaluation measure, the neighborhood decision error rate (NDER), for estimating classification complexity.
• To develop an effective feature selection algorithm applicable to both categorical and numerical features.
• To demonstrate the algorithm's superiority over existing methods in handling diverse datasets.

Main Methods:

• Introduction of a neighborhood rough-set model to partition sample sets into decision regions.
• Classification complexity estimation based on misclassified samples within boundary regions using neighborhood class probabilities.
• Implementation of a forward greedy search strategy to minimize the NDER and select optimal feature subsets.

Main Results:

• The NDER effectively estimates classification complexity across various feature subspaces.
• The proposed greedy algorithm successfully minimizes NDER, leading to reduced classification complexity.
• The feature selection method demonstrates efficacy on datasets with discrete, continuous, and mixed features.

Conclusions:

• The neighborhood decision error rate (NDER) provides a robust measure for feature evaluation and classification complexity estimation.
• The developed feature selection algorithm is effective for diverse data types, including mixed datasets.
• This approach offers a valuable tool for enhancing pattern recognition and machine learning performance.