Abstract
As regression techniques are increasingly developed and deployed in spectroscopic analysis, we have seen increasing implementation in many scientific and engineering disciplines. Therefore, ensuring their reliability and interpretability has become crucial. Theoretically, the characteristic responses occur in specific regions related to the chemical bonds of analytes, resulting in sparse and continuous feature structures. This paper proposes a Bayesian adaptive clustered prior learning (ACPL) method to capture and exploit such feature structures, thereby achieving state-of-the-art performance.
First, an unsupervised hierarchical clustering method is employed to identify the relationship between adjacent variables, clustering the spectral into a series of non-uniform blocks. Then, an initial prior will be arranged for each block. Since the importance of each block varies depending on the analytes, a Bayesian learning-based adaptive cluster block-prior inference model is introduced. This model considers intra-block variable interactions during the iterative process while adaptively penalizing blocks with lower contributions.
Extensive experiments on real datasets demonstrate that the model established by ACPL achieves superior performance, including state-of-the-art prediction accuracy and more interpretable results.
[Display omitted]
•Propose an adaptive Bayesian method for sparse spectroscopic regression.•The proposed method can capture and exploit feature structures within spectral.•The calibration model’s interpretability is enhanced compared to traditional chemometrics methods.•The proposed method yields impressive prediction improvement performance.