Scholarlink

<p>Our study addresses the challenges of understanding deep learning models applied to multivariate time-series data, particularly their perceived nature. We employ advanced post hoc explainability methods, TS-MULE and DeepSHAP on LSTM and Transformer models, revealing feature contribution values. This unravels key variables and time steps influencing predictions on Beijing Air Quality PM2.5 and Beijing Multi-Site Air Quality datasets. To tackle the evaluation challenge, we introduce a three-pronged approach. Firstly, perturbation analysis tests the reliability of feature importance scores. Secondly, model augmentation and retraining assess the impact of feature importance scores on model outcomes. Lastly, a comparative visual analysis of feature importance scores from LSTM and Transformer models is conducted. This comprehensive strategy enhances interpretability and provides a versatile, domain-agnostic framework for evaluating feature importance in diverse deep learning architectures. Moreover, it also contributes significantly to demystifying AI and promoting transparency.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Comprehensive Evaluation of Explainable AI for Multivariate Time-Series Regression Tasks

  • 期刊论文

  • 发表日期：

    2025年12月18日

查看全文
  • Veena More,
  • K. Ramesh,
  • Nadiyah Ahmed

摘要

Our study addresses the challenges of understanding deep learning models applied to multivariate time-series data, particularly their perceived nature. We employ advanced post hoc explainability methods, TS-MULE and DeepSHAP on LSTM and Transformer models, revealing feature contribution values. This unravels key variables and time steps influencing predictions on Beijing Air Quality PM2.5 and Beijing Multi-Site Air Quality datasets. To tackle the evaluation challenge, we introduce a three-pronged approach. Firstly, perturbation analysis tests the reliability of feature importance scores. Secondly, model augmentation and retraining assess the impact of feature importance scores on model outcomes. Lastly, a comparative visual analysis of feature importance scores from LSTM and Transformer models is conducted. This comprehensive strategy enhances interpretability and provides a versatile, domain-agnostic framework for evaluating feature importance in diverse deep learning architectures. Moreover, it also contributes significantly to demystifying AI and promoting transparency.