Abstract
Feature selection remains essential in machine learning, but traditional wrappers like Recursive Feature Elimination (RFE) suffer from quadratic complexity (O(d²)) and susceptibility to data leakage. We propose the Adaptive Interval Feature Selector (AIFS), a non-iterative wrapper-style selector grounded in aggregation functions. By mapping tree-based feature relevance scores into stability intervals across localized folds, AIFS constructs a global optimization sequence without sequential retraining, scaling near-linearly at O(M · d · N log N). Crucially, AIFS isolates feature selection from validation partitions to prevent data leakage. Empirical evaluation across 25 datasets demonstrates that AIFS achieves robust performance, yielding an average Balanced Accuracy of 0.8334 (outperforming standard wrappers), highly stable subset consistency (average Kuncheva Index of 0.6904), and resilience to 30 % missing data. Theoretical complexity scaling projects speedups of up to five orders of magnitude in ultra-high-dimensional spaces, successfully bridging the gap between filter efficiency and wrapper discriminative power.
Paper Type
Full Paper
DOI
10.62036/ISD.2026.40
Adaptive Interval Feature Selector Based on Aggregation Functions for Resilient and Computationally Optimal Learning
Feature selection remains essential in machine learning, but traditional wrappers like Recursive Feature Elimination (RFE) suffer from quadratic complexity (O(d²)) and susceptibility to data leakage. We propose the Adaptive Interval Feature Selector (AIFS), a non-iterative wrapper-style selector grounded in aggregation functions. By mapping tree-based feature relevance scores into stability intervals across localized folds, AIFS constructs a global optimization sequence without sequential retraining, scaling near-linearly at O(M · d · N log N). Crucially, AIFS isolates feature selection from validation partitions to prevent data leakage. Empirical evaluation across 25 datasets demonstrates that AIFS achieves robust performance, yielding an average Balanced Accuracy of 0.8334 (outperforming standard wrappers), highly stable subset consistency (average Kuncheva Index of 0.6904), and resilience to 30 % missing data. Theoretical complexity scaling projects speedups of up to five orders of magnitude in ultra-high-dimensional spaces, successfully bridging the gap between filter efficiency and wrapper discriminative power.
Recommended Citation
Wojtowicz, A. & Paja, W.(2026). Adaptive Interval Feature Selector Based on Aggregation Functions for Resilient and Computationally Optimal Learning. In M. Valenta, B. Mannová, R. Pergl, A. Przybylek, M. Lang, H. Linger, C. Schneider, N. Iivari, & E. Insfran (Eds.), Making ISD Sustainable: Reloaded with AI and Automation (ISD2026 Proceedings). Prague, Czech Republic: Czech Technical University in Prague. ISBN: 978-80-01-07585-2. https://doi.org/10.62036/ISD.2026.40