Abstract: Predictive business process monitoring focuses on the use of event logs to predict the future characteristics of running processes.Aiming at the shortcomings of the most existing methods of business process prediction,such as unable capture the long distance dependency of sequences and the sequence information only be employed in one direction,a next activity prediction method of business process based on XLNet was proposed.This method realized long range memory,and the attention mask was used to reconstruct the event sequence,so as to utilize the bidirectional information of the sequence.The evaluation results on 4 public datasets showed that the average accuracy of the proposed method was better than some popular prediction methods of next activity.When the log records were sufficient,this method could achieve high accuracy for the prediction result of next activity in business process,and provide a real-time decision-making basis for the business process management system.

Key words: business process, next activity prediction, deep learning, XLNet model

摘要： 预测性业务流程监控侧重于使用事件日志预测正在运行流程的未来特征,针对大多数现有业务流程预测方法的缺点,例如无法捕获序列的长距离依赖、只能单向利用序列信息,提出一种基于XLNet的业务流程下一活动预测方法。该方法实现了长程记忆,并采用注意力掩码重构事件序列,以利用序列的双向信息。通过在4个公开数据集上进行评估表明,该方法的平均准确率具有优越性,且在日志记录充分时,该方法对业务流程下一活动的预测准确率较高,可为业务流程管理系统提供实时的决策依据。

关键词: 业务流程实例, 下一活动预测, 深度学习, XLNet模型