Spectral Zones-Based\nSHAP/LIME: Enhancing Interpretability\nin Spectral Deep Learning Models Through Grouped Feature Analysis

Interpretability is just as important as accuracy when\nit comes\nto complex models, especially in the context of deep learning models.\nExplainable artificial intelligence (XAI) approaches have been developed\nto address this problem. The literature on XAI for spectroscopy mainly\nemphasizes independent feature analysis with limited application of\nzone analysis. Individual feature analysis methods, such as shapley\nadditive explanations (SHAP) and local interpretable model-agnostic\nexplanations (LIME), have limitations due to their dependence on perturbations.\nThese methods measure how AI models respond to sudden changes in the\nindividual feature values. While they can help identify the most impactful\nfeatures, the abrupt shifts introduced by replacing these values with\nzero or the expected ones may not accurately represent real-world\nscenarios. This can lead to mathematical and computational interpretations\nthat are neither physically realistic nor intuitive to humans. Our\nproposed method does not rely on individual disturbances. Instead,\nit targets “spectral zones” to directly estimate the\neffect of group disturbances on a trained model. Consequently, factors\nsuch as sample size, hyperparameter selection, and other training-related\nconsiderations are not the primary focus of the XAI methods. To achieve\nthis, we have developed a modified version of LIME and SHAP capable\nof performing group perturbations, enhancing explainability and realism\nwhile minimizing noise in the plots used for interpretability. Additionally,\nwe employed an efficient approach to calculate spectral zones for\ncomplex spectra with indistinct spectral boundaries. Users can also\ndefine the zones themselves using their domain-specific knowledge.