Abstract: Most existing deep learning-based methods for predicting the remaining time of business processes are mostly designed for a single execution instance,unable to perceive the influence of other concurrent instances in aspects such as resource competition.Additionally,there is a lack of in-depth exploration of instance embedding representation,resulting in considerable room for improvement in remaining time prediction.To address these deficiencies,a remaining time prediction method based on Bag-Attention for concurrent multi-instances was proposed,and a self-supervised learning-based instance embedding representation method was introduced to enhance the quality of instance representation.A method for constructing a concurrent instance dataset was presented,and event attributes were added based on event activities,enriching the input information for subsequent models;then,instance attribute embeddings were obtained through pre-training;finally,process instance attributes were encoded in the form of multi-channel convolution,and concurrent instances were integrated through the attention mechanism.Experimental results showed that the proposed method had effectively improved the accuracy of remaining time prediction compared to traditional methods.

Key words: business process management, prediction of remaining time, concurrent multiple instances, deep learning

摘要： 现有的基于深度学习的业务流程剩余时间预测方法大多针对单一执行实例构建预测模型,无法感知到同时执行的其他并发实例在资源竞争等方面的影响,并且对于实例嵌入表示缺少深入探究,导致现有方法的预测效果还有较大提升空间。针对现有方法的不足,提出了基于包注意力的并发多实例剩余时间预测方法,并引入了基于自监督学习的实例嵌入表示方法以提升实例表示的质量。首先,给出了并发实例数据集的构建方法,并在事件活动的基础上添加事件属性,丰富了后续模型的信息输入;之后,通过预训练的方式得到实例属性嵌入表示;最后,将流程实例属性以多通道卷积的形式进行编码,并通过注意力机制对并发实例进行融合。实验结果表明,所提方法与传统方法相比,有效提升了剩余时间预测的准确性。

关键词: 业务流程管理, 剩余执行时间预测, 并发多实例, 深度学习