Abstract: Business process prediction is an important research direction in Business Process Management (BPM) that aims to accurately forecast future behavioral events.It provides crucial support for downstream tasks in process behavior analysis.In the context of BPM research,most existing methods for process behavior analysis rely on black-box deep learning models,resulting in poor interpretability and an inability to provide explanations for prediction outcomes.To address this issue,a Hierarchical Bidirectional Encoder Representation from Transformers (BERT)-based Interpretability framework (HBI) was proposed,which realized interpretable ability from local to global (overall framework behavior) levels.Subsequently,the framework was subjected to interpretable analysis using hierarchical approaches,feature importance analysis and attention visualization to comprehend the internal workings and decision logic,thereby enhancing transparency.Experimental results on real-world event logs demonstrated that the HBI framework could achieve both higher prediction accuracy compared to state-of-the-art methods and interpretability.

Key words: business process management prediction, interpretability, deep learning, attention, event logs

摘要： 业务过程预测作为业务过程管理（BPM）的一个重要研究方向,用于准确预测未来的行为事件。它可以为过程行为分析方法的下游任务提供重要支持。面向BPM研究,大多数现有的过程行为分析方法都采用了基于黑箱的深度学习模型,导致可解释性较差,无法提供关于为什么要做出某个预测结果的解释。首先,提出一种基于BERT的层次化可解释性框架（HBI）,实现了从局部可解释扩展到全局（整体框架行为）可解释的能力。然后,基于层次化、特征重要性分析、注意力可视化的方法对框架进行可解释分析,理解内部运作过程和决策逻辑,提高透明度。最后,在真实的事件日志中的实验结果表明,相比最先进的研究方法,HBI框架既保证了高于基线的预测准确度,也确保了框架的可解释性。

关键词: 业务过程管理预测, 可解释性, 深度学习, 注意力, 事件日志