Spectral Zones-Based SHAP/LIME: Enhancing Interpretability in Spectral Deep Learning Models Through Grouped Feature...

Area of Science:

• Spectroscopy
• Artificial Intelligence
• Machine Learning

Background:

• Interpretability is crucial for complex deep learning models, particularly in spectroscopy.
• Existing Explainable AI (XAI) methods like SHAP and LIME often rely on individual feature perturbations, which can be unrealistic.
• Current XAI literature for spectroscopy primarily focuses on independent feature analysis, neglecting zone analysis.

Purpose of the Study:

• To develop a more realistic and interpretable XAI method for spectroscopy.
• To address the limitations of individual feature perturbation methods in XAI.
• To enhance the understanding of deep learning model behavior in spectroscopic analysis.

Main Methods:

• Developed modified versions of SHAP and LIME to perform group perturbations on spectral zones.
• Implemented an efficient approach for calculating spectral zones in complex spectra.
• Allowed users to define spectral zones based on domain expertise.

Main Results:

• The proposed method enhances the realism and interpretability of XAI in spectroscopy.
• Group perturbations on spectral zones provide more intuitive and physically meaningful explanations.
• Minimized noise in interpretability plots compared to individual feature analysis.

Conclusions:

• Analyzing spectral zones offers a more robust and realistic approach to XAI in spectroscopy.
• Modified LIME and SHAP methods effectively capture the impact of group disturbances.
• The developed XAI technique improves the trustworthiness and applicability of deep learning models in spectroscopic applications.