Small-sample molecular data — such as omics profiles from limited patient cohorts — make feature selection difficult because many candidate variables are measured on few examples. A new preprint proposes ARISE (Adaptive Residual-Informed Stability Ensemble), a method that combines multiple relevance measures with stability-aware, redundancy-residual, and multiclass-coverage mechanisms. The algorithm automatically weights seven percentile-normalized relevance components using 15 predefined profiles, tuned by nested inner cross-validation.
The authors evaluated ARISE on five molecular datasets, eight feature-set sizes, and three fixed classifiers: k-nearest neighbours, support vector machine, and random forest. Across 210,000 held-out assessments, ARISE ranked first in all 15 dataset-metric combinations. Equal-dataset means were 0.793 for balanced accuracy, 0.776 for macro-F1, and 0.725 for Cohen's kappa, exceeding the strongest aggregate comparator by 0.022, 0.023, and 0.028, respectively.
The results suggest that ARISE's adaptive weighting helps maintain competitive performance on small and compact feature sets, although the optimal number of features differed by dataset. The authors frame the method as a transparent alternative to black-box selection pipelines, and propose further evaluation in small-sample molecular classification tasks. As a preprint posted on arXiv, it has not yet undergone journal peer review.