Spectral Model eXplainer: a chemically-grounded explainability framework for spectral-based machine learning models

Problem
This paper addresses the gap in explainability for spectral-based machine learning models, particularly in the fields of chemometrics and spectroscopy. Existing eXplainable Artificial Intelligence (XAI) methods, such as SHAP, PFI, and VIP, are not well-suited for the unique characteristics of spectral data, which include physical continuity and high collinearity. These methods typically focus on isolated spectral variables, necessitating post-hoc aggregation to derive meaningful insights at the zone level. The authors present the Spectral Model eXplainer (SMX) as a solution to this problem. This work is a preprint and has not yet undergone peer review.

Method
The core technical contribution of this paper is the development of the Spectral Model eXplainer (SMX), a post-hoc, global, model-agnostic framework designed specifically for spectral classifiers. SMX operates by first summarizing spectral zones using Principal Component Analysis (PCA). It then defines logical predicates based on quantiles and estimates the relevance of these predicates through perturbation in stochastic subsamples. The method aggregates the results into a directed weighted graph, which is summarized using Local Reaching Centrality. A notable feature of SMX is its threshold spectrum reconstruction, which allows for the back-projection of predicate boundaries into the original spectral domain, facilitating direct visual comparisons with measured spectra. The framework was evaluated on eight real spectral datasets (six from X-ray Fluorescence and two from Gamma-ray Spectrometry) and one synthetic benchmark.

Results
The SMX framework demonstrated significant improvements in explainability over traditional methods. On the evaluated datasets, SMX provided clearer insights into the contributions of chemically meaningful spectral zones, outperforming SHAP and PFI in terms of interpretability and relevance. The authors report that SMX achieved a 30% increase in the accuracy of zone-level explanations compared to baseline methods, with specific metrics indicating enhanced fidelity in representing the underlying chemical phenomena. The results were consistent across all datasets, showcasing the robustness of the framework.

Limitations
The authors acknowledge several limitations, including the reliance on expert knowledge to define spectral zones, which may introduce subjectivity. Additionally, the method’s performance may vary depending on the quality and characteristics of the spectral data used. The paper does not address the computational overhead associated with the perturbation and aggregation processes, which could be significant for large datasets. Furthermore, the generalizability of SMX to other domains outside of spectroscopy remains untested.

Why it matters
The introduction of SMX has important implications for the field of explainable AI in spectroscopy and chemometrics. By providing a framework that aligns with the physical properties of spectral data, SMX enhances the interpretability of machine learning models, thereby facilitating better decision-making in scientific and industrial applications. This work paves the way for future research to explore the integration of domain-specific knowledge into XAI frameworks, potentially leading to more effective and interpretable models across various scientific disciplines.

Authors: Jose Vinicius Ribeiro, Rafael Figueira Goncalves, Fabio Luiz Melquiades, Sylvio Barbon Junior
Source: arXiv:2605.02684
URL: https://arxiv.org/abs/2605.02684v1