Unified Time Series Framework for Explainable Artificial Intelligence

• Ano de defesa: 2025
• Autor(a) principal: Bragança, Hendrio Luis de Souza
• Outros Autores: http://lattes.cnpq.br/7375456368369943, https://orcid.org/0000-0003-1479-1707
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Universidade Federal do Amazonas; Instituto de Computação; Brasil; UFAM; Programa de Pós-graduação em Informática
• Palavras-chave em Português: CIENCIAS EXATAS E DA TERRA: CIENCIA DA COMPUTACAO; Explainable Artificial Intelligence (XAI); Time-Series data; Machine Learning
• Link de acesso: https://tede.ufam.edu.br/handle/tede/11020

The increasing complexity of machine learning (ML) models has made their decision-making processes difficult to interpret, posing a critical challenge in high-stakes domains where trust and transparency are essential. Although Explainable Artificial Intelligence (XAI) methods aim to address this issue, most existing techniques face limitations when applied directly to time series data due to its sequential and contextual nature. In this work, we present the Unified Time Series Framework for Explainable Artificial Intelligence (UTS-XAI), which integrates a standard time series classification pipeline with explainability capabilities and domain-specific evaluation tools. The framework is compatible with multiple explainability methods, such as SHAP, LIME, and Saliency Maps, and supports their systematic evaluation through adapted versions of widely used XAI metrics (faithfulness, robustness, sensitivity, and stability) reinterpreted for temporal data. These metrics are combined with time series–specific similarity and distance measures such as MSE, MAE, and DTW to quantify explanation quality. We also introduce Global Interpretable Clustering (GIC), a visualization technique designed to assess the consistency of feature attributions across explainers and models. Experiments conducted on three real-world cardiac arrhythmia datasets (MITBIH, SVDB, INCART), using three ML architectures (XGBoost, DeepConvLSTM, and FCN), show that SHAP provides more faithful and stable explanations, while LIME and Saliency Maps exhibit greater sensitivity to noise and perturbations. These results highlight that accuracy alone is not sufficient in time series modeling without robust interpretability. By embedding explainability into the model development lifecycle, UTS-XAI sets a new standard for interpretable and trustworthy AI in temporal data analysis.