Abstract
In this paper, we propose a knowledge-informed spectroscopic regression method named burst-sparsity learning (BSL) to address limitations in interpretability and consistency analysis. The concept of burst-sparsity (BS) refers to the distribution of chemically relevant structures inspired by spectral response mechanisms, characterized by significant variables that are sparse and occur in clusters. First, we formulate spectroscopic regression as a sparse recovery problem using the sparse Bayesian learning (SBL) model, which leverages the flexibility of SBL to provide an accurate sparse representation and allows for the integration of prior knowledge. Second, since the BS structure is unavailable, an enhanced non-uniform pattern-coupled (PC) prior was developed to capture more BS structures by considering adjacent coefficients. Extensive experiments are conducted to verify the efficacy of the BSL method. The results show that the BSL enhances the prediction performance in term of RMSEP and Rp2 across various spectroscopic techniques and dataset scales, highlighting its impressive potential for real-world applications. In additional, the deep combination of domain knowledge into machine learning provides deeper insights into how chemically relevant features contribute to the model’s predictions.
•Propose a novel Burst-sparsity learning (BSL) framework for spectroscopic regression.•Domain knowledge and spectral mechanism were integrated with machine learning.•A non-uniform pattern-coupled prior was proposed to capture and exploit feature structure.•The BSL shows advanced performance in accurate; consistency and interpretability.•The BSL provides deeper insights into how spectra variable contributes to prediction.