Scholarlink

<p>Deep learning models promise a lot for predicting job cycle times in manufacturing but they often fall short when working with the limited data sets typical of real-world factories. To tackle this generalization problem, we have developed a post-hoc correction framework that uses SHAP explainability to make smarter adjustments. Instead of applying a blanket fix our method first pinpoints which new jobs are likely to have inaccurate predictions. Then for only those jobs it uses insights from SHAP to fine-tune the most critical input attributes before feeding them back into the original unchanged deep neural network (DNN). When we tested this on a real 300-job data set from a wafer fab, the results were striking the framework slashed the Root Mean Square Error (RMSE) by 22.5%. This correction brought the DNN’s performance to a level that was better than or just as good as standard benchmarks like Random Forests showing it’s a practical way to get more out of deep learning models even when data is scarce.</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

Explainability driven post-hoc correction enhancing deep neural network performance for job cycle time prediction

  • 期刊论文

  • 发表日期：

    2025年11月30日

查看全文
  • Yu-Cheng Wang

摘要

Deep learning models promise a lot for predicting job cycle times in manufacturing but they often fall short when working with the limited data sets typical of real-world factories. To tackle this generalization problem, we have developed a post-hoc correction framework that uses SHAP explainability to make smarter adjustments. Instead of applying a blanket fix our method first pinpoints which new jobs are likely to have inaccurate predictions. Then for only those jobs it uses insights from SHAP to fine-tune the most critical input attributes before feeding them back into the original unchanged deep neural network (DNN). When we tested this on a real 300-job data set from a wafer fab, the results were striking the framework slashed the Root Mean Square Error (RMSE) by 22.5%. This correction brought the DNN’s performance to a level that was better than or just as good as standard benchmarks like Random Forests showing it’s a practical way to get more out of deep learning models even when data is scarce.