Scholarlink

<p>Recent advances in Machine Learning (ML) have shown significant potential to support proactive decision-making in industrial processes. However, the opaque nature of ML models can hinder users’ adoption, especially in high-stakes environments. eXplainable AI (XAI) aims to address this, but traditional approaches often neglect user needs. Human-Centered XAI (HC-XAI) emphasizes designing explanations that align with users’ goals and mental models. This study presents a human-centered XAI solution for operators managing the complex delignification stage of the pulp process. Through a multiyear, iterative design process, we developed and evaluated a transformer-based time series forecasting dashboard with eight professional operators. Key findings highlight the value of uncertainty bands, historical trend comparisons, and interactive features. This work contributes to HC-XAI by designing a functional dashboard for real world use cases, generating empirical insights and design implications for developing ML explanations that align with domain experts’ mental models in high-stake industrial environments.</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

Human-Centered Explainable AI for Time Series Forecasting: A Case Study in the Pulp and Paper Industry

  • 期刊论文

  • 发表日期：

    2026年3月6日

查看全文
  • Yanqing Zhang,
  • Hugo Wärnberg,
  • Ruben Huehnerbein,
  • Nilavra Bhattacharya,
  • Emmanuel Brorsson,
  • Andreas Darnell,
  • Fabian Buelow

摘要

Recent advances in Machine Learning (ML) have shown significant potential to support proactive decision-making in industrial processes. However, the opaque nature of ML models can hinder users’ adoption, especially in high-stakes environments. eXplainable AI (XAI) aims to address this, but traditional approaches often neglect user needs. Human-Centered XAI (HC-XAI) emphasizes designing explanations that align with users’ goals and mental models. This study presents a human-centered XAI solution for operators managing the complex delignification stage of the pulp process. Through a multiyear, iterative design process, we developed and evaluated a transformer-based time series forecasting dashboard with eight professional operators. Key findings highlight the value of uncertainty bands, historical trend comparisons, and interactive features. This work contributes to HC-XAI by designing a functional dashboard for real world use cases, generating empirical insights and design implications for developing ML explanations that align with domain experts’ mental models in high-stake industrial environments.