Abstract: Business process remaining time prediction can effectively help enterprises deal with business risks.The existing deep learning-based remaining time prediction methods suffer from single event vector representation,inability to dynamically represent the trajectory itself,low model computational efficiency and insufficient long sequence modeling capability.To address these issues,a BiLSTM-based event vector representation method and a sparse attention mechanism-based remaining time prediction model were proposed.A remaining time prediction model was built based on Informer encoder,where the feature sampling layer of the Informer encoder was improved from a regular convolution to a dilated causal convolution to enhance performance.Then a BiLSTM-based dynamic event vector representation method was proposed,which dynamically represented the events in different trajectory prefixes as vector to improve the remaining time prediction effect.Experiments on 7 public event log datasets showed that the proposed method could effectively improve the accuracy of remaining time prediction,with an average improvement of 30% in prediction accuracy compared to existing methods.

Key words: remaining time prediction, process mining, deep learning, Informer model

摘要： 业务流程剩余时间预测可以有效地帮助企业应对业务风险。现有的基于深度学习的预测方法存在事件的静态表征难以捕捉轨迹动态变化、长序列建模能力不足的问题。针对上述问题,提出一种基于BiLSTM的事件向量表示方法和基于稀疏注意力机制的剩余时间预测模型。首先以Informer编码器为基础构建剩余时间预测模型,将编码器中特征采样层的普通卷积改进为扩张因果卷积,以提升性能。其次基于双向长短期记忆(BiLSTM)的动态事件向量表示法,实现了对不同轨迹中的事件进行动态的向量表示,达到提升剩余时间预测效果的目的。经过在7个公开事件日志数据集上的实验表明,该方法可以有效提升剩余时间预测的精度,与已有的方法在预测精度上平均提升了30%。

关键词: 剩余时间预测, 过程挖掘, 深度学习, Informer模型