Scholarlink

<p>Predictive Business Process Monitoring (PBPM) aims to forecast future events in ongoing business processes, particularly the next activity and its timestamp. Although deep learning methods such as LSTM and CNN have achieved notable success in process prediction tasks, existing models often suffer from instability and limited generalization, primarily caused by sparse and imbalanced event logs. To address these limitations, an Enhanced Bidirectional Generative Adversarial Network (E-BiGAN) is proposed in this paper, that realizes multi-task prediction of the next activity and its timestamp. E-BiGAN extends the standard BiGAN by integrating a context-aware LSTM-based generator for sequence-conditioned outputs, employing Wasserstein GAN with Gradient Penalty (WGAN-GP) to ensure training stability, and utilizing a joint optimization strategy that balances classification and regression objectives. Meanwhile, a dual LSTM encoder improves latent representations, and a scalable preprocessing pipeline ensures adaptability across diverse log formats. Experiments on seven real-life event logs demonstrate that E-BiGAN outperforms baseline methods, enhancing predictive accuracy and robustness across various domains.</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

Enhanced BiGAN for multi-task learning in business process prediction: next activity and timestamp prediction tasks

  • 期刊论文

  • 发表日期：

    2025年10月17日

查看全文
  • Hang Li,
  • Lili Wang

摘要

Predictive Business Process Monitoring (PBPM) aims to forecast future events in ongoing business processes, particularly the next activity and its timestamp. Although deep learning methods such as LSTM and CNN have achieved notable success in process prediction tasks, existing models often suffer from instability and limited generalization, primarily caused by sparse and imbalanced event logs. To address these limitations, an Enhanced Bidirectional Generative Adversarial Network (E-BiGAN) is proposed in this paper, that realizes multi-task prediction of the next activity and its timestamp. E-BiGAN extends the standard BiGAN by integrating a context-aware LSTM-based generator for sequence-conditioned outputs, employing Wasserstein GAN with Gradient Penalty (WGAN-GP) to ensure training stability, and utilizing a joint optimization strategy that balances classification and regression objectives. Meanwhile, a dual LSTM encoder improves latent representations, and a scalable preprocessing pipeline ensures adaptability across diverse log formats. Experiments on seven real-life event logs demonstrate that E-BiGAN outperforms baseline methods, enhancing predictive accuracy and robustness across various domains.