Bayesian optimization for interval selection in PLS models

Abstract

We propose a novel Bayesian optimization framework for interval selection in Partial Least Squares (PLS) regression. Unlike traditional iPLS variants that rely on fixed or grid-based intervals, our approach adaptively searches over the discrete space of interval positions of a pre-defined width using a Gaussian Process surrogate model and an acquisition function. This enables the selection of one or more informative spectral regions without exhaustive enumeration or manual tuning. Through synthetic and real-world spectroscopic datasets, we demonstrate that the proposed method consistently identifies chemically relevant intervals, reduces model complexity, and improves predictive accuracy compared to full-spectrum PLS and stepwise interval selection techniques. A Monte Carlo study further confirms the robustness and convergence of the algorithm across varying signal complexities and uncertainty levels. This flexible, data-efficient approach offers an interpretable and computationally scalable alternative for chemometric applications.