Abstract: Remaining time prediction can effectively enhance the risk response ability of enterprises.The existing methods have problems such as insufficient corpus in trajectory characterization,difficulty in capturing key information,limited and universal deep learning models applied,and the need to train multiple models based on different lengths.To address the above issues,a method for predicting absolute remaining time based on prefix trajectory representation learning method and attention mechanism was proposed.A prefix trajectory representation learning method was designed to obtain the representation vector.Then,the PTr-Transformer model was proposed by combining with the attention mechanism.Finally,the model was tested on 5 real event logs,and the results showed that it could effectively improve the remaining time prediction accuracy for large-scale datasets,with a maximum improvement of 8.3%.

Key words: remaining time prediction, business process management, pre-trace, attention mechanism, represent learningremaining time prediction, business process management, pre-trace, attention mechanism, represent learning

摘要： 剩余时间预测可以提升企业的风险应对能力,现有的预测方法存在轨迹刻画中语料库不丰富难以捕捉关键信息、应用的深度学习模型单一且通用性不足以及需要根据不同长度训练多个模型等问题。针对上述问题,提出一种基于前缀轨迹表示学习方法和注意力机制的绝对剩余时间预测模型。首先,设计一种前缀轨迹表示学习方法获取表示向量,然后结合注意力机制提出PTr-Transformer模型。最后,该模型在5个真实事件日志中进行实验,结果表明针对大规模数据集可以有效提升剩余时间预测精度,最高可提升8.3%。

关键词: 剩余时间预测, 业务流程管理, 前缀轨迹, 注意力机制, 表示学习