<p>The growing complexity of machine learning (ML) models has made their decision-making processes increasingly opaque, posing a critical challenge in high-stakes domains where trust, transparency, and accountability are essential. While Explainable Artificial Intelligence (XAI) techniques attempt to mitigate this issue, most existing approaches exhibit important limitations when applied to time series data due to its inherently sequential, dynamic, and context-dependent structure. In this work, we introduce the Unified Time Series Classification Framework for Explainable Artificial Intelligence (UTS-XAI), which integrates a standard classification pipeline with comprehensive, time-aware XAI evaluation metrics, such as faithfulness, robustness, sensitivity, stability, and localization. The framework supports multiple explanation methods, such as SHAP, LIME, and Saliency Maps, and enables their systematic comparison using adapted versions of widely used XAI metrics (faithfulness, robustness, sensitivity, and stability) reinterpreted for temporal data. We further propose Global Interpretable Clustering (GIC), a visualization technique designed to evaluate the consistency of feature attributions across different explainers and model architectures. Experiments on the MIT-BIH, SVDB, and INCART datasets show that DeepConvLSTM achieved robust baseline performance, maintaining an F1-score of 93.5% on the external INCART test set. Our quantitative analysis reveals that globally consistent methods such as SHAP significantly outperform localized techniques such as LIME and Saliency Maps. SHAP exhibits extreme faithfulness; ablating its top 10% of identified features collapses models’ true positive detection rates from <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(>85\%\)</EquationSource> </InlineEquation> to <InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(<27\%\)</EquationSource> </InlineEquation>. Furthermore, SHAP yields highly stable explanations across diverse deep architectures (median Dynamic Time Warping distance  1.5) and maintains superior structural robustness under noise (SSIM  0.95). Overall, our findings demonstrate that accuracy alone is insufficient in time series modeling without robust and reliable interpretability. By embedding explainability throughout the entire model development lifecycle, UTS-XAI establishes a unified foundation for interpretable, transparent, and trustworthy AI in temporal data analysis.</p>

Unified time series classification framework for explainable artificial intelligence

期刊论文
出版模式：

开放获取
发表日期：

2026年4月28日

查看全文

Hendrio Bragança,
Eduardo Souto

摘要

The growing complexity of machine learning (ML) models has made their decision-making processes increasingly opaque, posing a critical challenge in high-stakes domains where trust, transparency, and accountability are essential. While Explainable Artificial Intelligence (XAI) techniques attempt to mitigate this issue, most existing approaches exhibit important limitations when applied to time series data due to its inherently sequential, dynamic, and context-dependent structure. In this work, we introduce the Unified Time Series Classification Framework for Explainable Artificial Intelligence (UTS-XAI), which integrates a standard classification pipeline with comprehensive, time-aware XAI evaluation metrics, such as faithfulness, robustness, sensitivity, stability, and localization. The framework supports multiple explanation methods, such as SHAP, LIME, and Saliency Maps, and enables their systematic comparison using adapted versions of widely used XAI metrics (faithfulness, robustness, sensitivity, and stability) reinterpreted for temporal data. We further propose Global Interpretable Clustering (GIC), a visualization technique designed to evaluate the consistency of feature attributions across different explainers and model architectures. Experiments on the MIT-BIH, SVDB, and INCART datasets show that DeepConvLSTM achieved robust baseline performance, maintaining an F1-score of 93.5% on the external INCART test set. Our quantitative analysis reveals that globally consistent methods such as SHAP significantly outperform localized techniques such as LIME and Saliency Maps. SHAP exhibits extreme faithfulness; ablating its top 10% of identified features collapses models’ true positive detection rates from \(>85\%\) to \(<27\%\) . Furthermore, SHAP yields highly stable explanations across diverse deep architectures (median Dynamic Time Warping distance 1.5) and maintains superior structural robustness under noise (SSIM 0.95). Overall, our findings demonstrate that accuracy alone is insufficient in time series modeling without robust and reliable interpretability. By embedding explainability throughout the entire model development lifecycle, UTS-XAI establishes a unified foundation for interpretable, transparent, and trustworthy AI in temporal data analysis.