Knowledge discovery by accuracy maximization.

Area of Science:

• Computational Biology
• Data Science
• Machine Learning

Background:

• Feature extraction from high-dimensional and noisy data is challenging.
• Existing data mining methods often lack robust validation procedures.
• Unsupervised and semi-supervised learning algorithms are crucial for knowledge discovery.

Purpose of the Study:

• To introduce KODAMA (knowledge discovery by accuracy maximization), a novel algorithm for feature extraction.
• To highlight KODAMA's integrated cross-validation for enhanced result robustness.
• To demonstrate KODAMA's effectiveness across various scientific domains.

Main Methods:

• KODAMA employs an unsupervised and semi-supervised learning approach.
• It utilizes a Monte Carlo procedure to maximize cross-validated predictive accuracy for manifold topology discovery.
• The algorithm integrates a validation procedure driven by cross-validation results.

Main Results:

• KODAMA demonstrates superior performance on gene expression and metabolomics datasets compared to existing methods.
• The algorithm successfully identified unexpected features in an astronomical dataset.
• Analysis of State of the Union speeches revealed a distinct linguistic transition during the Reagan presidency.

Conclusions:

• KODAMA offers a robust and effective solution for feature extraction from complex datasets.
• Its integrated validation enhances the reliability of discovered patterns.
• The algorithm shows broad applicability in fields ranging from biology to political science.