Interval retention optimization (IRO): An efficient feature selection method for expanding spectral datasets

Effective feature selection is crucial for large-scale near-infrared (NIR) spectroscopy, yet existing algorithms face a trade-off between accuracy and efficiency. This trade-off arises from the search strategy: sequential methods are efficient but often lack generalization, while global methods capture feature interactions but incur high computational costs due to repeated retraining. To address these limitations, we propose Interval Retention Optimization (IRO), a framework that reformulates feature selection as a continuous allocation of retention rates across wavelength intervals. Guided by global importance measures and optimized with Bayesian search, IRO leverages a mask-based perturbation strategy to evaluate candidate subsets directly on a pre-trained model, thereby eliminating retraining and significantly boosting efficiency. Experimental results demonstrate that IRO can achieve improved prediction accuracy and computational efficiency, reducing RMSEP by up to 9.10 %, improving RMSECV and R² by up to 5.51 % and 15.20 %, respectively, and accelerating computation by up to 87.54 %. These results highlight IRO as a scalable and practical solution for spectral feature selection in complex NIR applications.