RRCT

RRCT optimizes selection of features in bioinformatics datasets by balancing relevance, redundancy, and complementarity to identify informative variables for predictive modeling.

Key Features:

• Heuristic feature-selection algorithm: RRCT implements a heuristic procedure that trades off relevance, redundancy, and complementarity when selecting feature subsets.
• Information-theoretic framework: Quantifies relevance, redundancy, and complementarity within an information-theoretic framework.
• Rank correlation coefficients: Uses rank correlation coefficients to assess association strength between each feature and the response variable and between pairs of features.
• Partial correlation coefficients: Evaluates complementarity among features using partial correlation coefficients.
• Benchmarking: Was empirically benchmarked against 19 other feature-selection algorithms across four synthetic and eight real-world datasets.
• Evaluation criteria: Assesses recovery of true underlying relevant features and out-of-sample predictive performance in classification tasks.
• Classifier assessment: Evaluates selected feature sets as inputs to random forest models for binary and multi-class classification.

Scientific Applications:

• Feature selection for predictive modeling: Identifies informative variables for predictive models in complex biological datasets.
• Dimensionality reduction for classification: Reduces feature dimensionality for binary and multi-class classification tasks.
• Identification of true relevant features: Recovers true underlying relevant features in synthetic and real-world datasets.
• Method comparison and benchmarking: Serves as a method for comparative benchmarking against other feature-selection algorithms.

Methodology:

Uses an information-theoretic framework that quantifies relevance and redundancy via rank correlation coefficients (feature–response and feature–feature) and evaluates complementarity via partial correlation coefficients; implemented as a heuristic feature-selection algorithm and empirically benchmarked against 19 other algorithms on four synthetic and eight real-world datasets with evaluation of feature recovery and out-of-sample binary and multi-class classification performance using random forest.