Recursive computed ABC (cABC) analysis as a precise method for reducing machine learning based feature sets to their...

Area of Science:

• Machine Learning
• Data Science
• Bioinformatics

Background:

• Feature selection is crucial in machine learning, often facing challenges with right-skewed feature importance distributions.
• Identifying the informative minimum of features is essential for efficient data analysis and model performance.

Purpose of the Study:

• To propose and evaluate a numerically precise method, computed ABC analysis (cABC), for reducing feature sets.
• To address skewed feature importance distributions by partitioning items into "A", "B", and "C" subsets.
• To demonstrate the effectiveness of recursive cABC analysis in feature selection for large datasets.

Main Methods:

• Computed ABC analysis (cABC) was employed, partitioning non-negative numerical items based on ABC curves and Lorenz curve relationships.
• Recursive application of cABC to subset "A" was performed for further refinement.
• Experiments were conducted on a generic image dataset and three biomedical datasets (lipidomics, genomics).

Main Results:

• Recursive cABC analysis effectively reduced data dimensions while preserving relevant information.
• Feature sets were reduced to 10% or less of original variables, maintaining accurate classification.
• The method successfully filtered out irrelevant variables and directed feature selection to class-relevant information.

Conclusions:

• Recursive cABC analysis offers a computationally precise method for minimizing feature sets to the most relevant items.
• The method provides precise criteria for stopping the reduction process, enhancing data understanding.
• The cABC method is available as a Python package, facilitating its application in machine learning and data analysis.