Explainable predictive modeling for limited spectral data

Feature selection of high-dimensional labeled data with limited observations is critical for making powerful predictive modeling accessible, scalable, and interpretable domain experts. Spectroscopy data, which records the interaction between matter electromagnetic radiation, particularly holds a lot information in single sample. Since acquiring such complex task, it crucial to exploit best analytical tools extract necessary information. In this paper, we investigate most commonly used feature techniques introduce applying recent explainable AI interpret prediction outcomes spectral data. Interpretation outcome beneficial experts as ensures transparency faithfulness ML models knowledge. Due instrument resolution limitations, pinpointing important regions spectroscopy creates pathway optimize collection process through miniaturization spectrometer device. Reducing device size power therefore cost requirement real-world deployment sensor-to-prediction system whole. Furthermore, consider wide range machine learning that have been proven be successful Cetane Number fuels. We specifically design three different scenarios ensure evaluation robust real-time practice developed methodologies uncover hidden effect noise sources on final outcome. The performed both full model reduced using real dataset. Finally, propose correctness metric assess conformance selected subset features expertise. As result, Support Vector Regression yields better accuracy generalization leads less computationally more efficient than Neural Network. More importantly, from original deploying complex, models.