业务流程剩余时间预测对于预防流程超时风险有着重要意义。深度学习模型的应用显著提高了剩余时间预测的准确率,但由于输入信息单一,难以发挥模型自身优势;此外,有效的活动编码包含丰富的上下文关系,有助于提升预测效果。针对上述问题,本研究提出一种基于Transformer的业务流程剩余时间预测方法。首先,基于Transformer网络架构构建剩余时间预测模型,选取活动和时间属性作为模型的输入特征。然后,为捕获业务流程活动间的时序关系,选取4种主流自然语言处理领域的编码方式训练活动的表示向量,以编码方式与剩余时间预测任务的匹配性、编码维度两个角度对4种编码方式进行评估。最后,采用6个真实事件日志进行实验测试,结果表明本研究所提方法与已有方法相比预测准确率显著提升,采用连续词袋模型方式训练的活动编码取得了更好的预测效果,并且基于不同规模的事件日志推荐了适用的编码维度。

The prediction of business process remaining time is of great significance to the prevention of process timeout risk. The application of the deep learning model improves the accuracy of the remaining time prediction. However, the advantages of the model are not fully utilized due to the single input information. In addition, an effective activity encoding approach can contain rich context to improve the prediction accuracy. In view of the above problems, this paper proposed a business process remaining time prediction approach based on Transformer. The Transformer network was first used to construct the remaining time prediction model, and the activity and time attributes are selected as the input features. Then, in order to capture the temporal relationship between business process activities, four mainstream natural language processing fields were selected to train the representation vectors of activities. The four approaches were evaluated from the perspectives of encoding adaptability and encoding dimension. Finally, six real event logs were used for experiments. The results show that the proposed approach, compared with the existing approaches, can significantly improve the prediction accuracy, that the activity encoding of continuous bag-of-word model (CBOW) achieves better prediction results, and that the encoding dimensions are recommended based on event logs of different scales.