Scholarlink

Predicting Work-in-Progress (WiP) is crucial for efficient process management. While approaches based on deep neural networks have become popular in this context, they often struggle in data-scarce environments, common in new or small-scale processes. This paper addresses this challenge by introducing a meta-learning approach that leverages Model-Agnostic Meta-Learning (MAML) combined with a Temporal Convolutional Network (TCN) architecture. Our method utilizes a comprehensive framework that transforms event logs into OHLC (Open, High, Low, Close)-like time series data and generates diverse training tasks through window and meta-window techniques. We demonstrate the effectiveness of our approach on small real-world datasets, where it achieves significant improvements in prediction accuracy compared to a baseline Long Short-Term Memory (LSTM) model, outperforming it in 93 out of 100 experiments. Our findings highlight the potential of meta-learning to enhance the robustness and adaptability of deep learning models for critical tasks in predictive process monitoring domains like WiP prediction, enabling accurate forecasting in environments with limited historical data.

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

Using Meta-Learning to Predict Work-in-Progress: An Approach for Small Datasets

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Yousef Mehrdad Bibalan,
  • Behrouz Far,
  • Mohammad Moshirpour,
  • Bahareh Ghiyasian

摘要

Predicting Work-in-Progress (WiP) is crucial for efficient process management. While approaches based on deep neural networks have become popular in this context, they often struggle in data-scarce environments, common in new or small-scale processes. This paper addresses this challenge by introducing a meta-learning approach that leverages Model-Agnostic Meta-Learning (MAML) combined with a Temporal Convolutional Network (TCN) architecture. Our method utilizes a comprehensive framework that transforms event logs into OHLC (Open, High, Low, Close)-like time series data and generates diverse training tasks through window and meta-window techniques. We demonstrate the effectiveness of our approach on small real-world datasets, where it achieves significant improvements in prediction accuracy compared to a baseline Long Short-Term Memory (LSTM) model, outperforming it in 93 out of 100 experiments. Our findings highlight the potential of meta-learning to enhance the robustness and adaptability of deep learning models for critical tasks in predictive process monitoring domains like WiP prediction, enabling accurate forecasting in environments with limited historical data.