Multivariate Optimization of k for k-Nearest-Neighbor Feature Selection With Dichotomous Outcomes: Complex...

Area of Science:

• Machine Learning
• Bioinformatics
• Computational Biology

Background:

• Nearest-neighbor feature selection is crucial but sensitive to data characteristics like sample size, feature count, statistical effects, and class imbalance.
• Existing methods, including Nearest-neighbor Projected-Distance Regression (NPDR), often do not adequately address class imbalance.
• Effective feature selection is vital for accurate biological data analysis and disease-related gene discovery.

Purpose of the Study:

• To develop and evaluate novel fixed-k strategies for nearest-neighbor feature selection that explicitly handle class imbalance.
• To improve the performance of NPDR and Relief-based algorithms on imbalanced datasets.
• To compare the efficacy of these new methods against established techniques like random forest and ridge regression.

Main Methods:

• Introduced minority-class-k for NPDR, parameterizing fixed-k by minority class size.
• Developed hit-miss-k, a class-adaptive fixed-k for Relief-based algorithms.
• Presented variable-wise optimized k (VWOK) and principal components analysis-derived k (kPCA) for imbalance-adaptive optimization.
• Utilized simulated data and consensus-nested cross-validation (cnCV) for rigorous performance evaluation.

Main Results:

• The proposed methods significantly enhanced feature detection across various nearest-neighbor scoring metrics.
• Novel fixed-k strategies demonstrated superior performance compared to random forest and ridge regression.
• Application to Major Depressive Disorder RNASeq data revealed NPDR with minority-class-k identified highly relevant genes for the condition.

Conclusions:

• The developed fixed-k methods effectively address class imbalance in nearest-neighbor feature selection.
• These adaptive strategies offer improved accuracy and relevance in feature discovery, particularly for imbalanced biological datasets.
• NPDR with minority-class-k shows potential as a reliable tool for identifying biologically significant genes in complex disease studies.