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

Area of Science:

• Analytical Chemistry
• Spectroscopy
• Chemometrics

Background:

• Effective feature selection is critical for large-scale near-infrared (NIR) spectroscopy.
• Existing algorithms present a trade-off between prediction accuracy and computational efficiency.
• Sequential methods are efficient but generalize poorly, while global methods are computationally expensive due to retraining.

Purpose of the Study:

• To introduce Interval Retention Optimization (IRO), a new framework for spectral feature selection.
• To address the accuracy-efficiency trade-off in NIR spectroscopy.
• To enhance scalability and practicality for complex NIR applications.

Main Methods:

• Reformulating feature selection as continuous retention rate allocation across wavelength intervals.
• Utilizing global importance measures and Bayesian optimization.
• Employing a mask-based perturbation strategy to evaluate feature subsets on a pre-trained model, avoiding retraining.

Main Results:

• IRO achieved improved prediction accuracy, reducing RMSEP by up to 9.10%, RMSECV by 5.51%, and improving R² by 15.20%.
• Significant computational efficiency gains were observed, with acceleration up to 87.54%.
• The proposed method demonstrated superior performance compared to existing approaches.

Conclusions:

• IRO offers a scalable and practical solution for spectral feature selection in NIR spectroscopy.
• The framework effectively balances prediction accuracy and computational efficiency.
• IRO represents a significant advancement for analyzing complex spectral data.