摘要： 现有的剩余时间预测方法仅关注对剩余时间预测任务起决定性作用的时间特征信息，并未考虑空间特征信息以及异质事件日志对预测任务的影响，导致预测准确度降低。提出基于轨迹聚类的剩余时间预测方法。将不同轨迹间的相似度作为距离度量，通过对事件日志中不同长度的轨迹进行聚类，以降低事件日志复杂度并细化结构。针对业务流程剩余时间预测任务，结合卷积神经网络与准循环神经网络，同时引入双向机制与注意力机制，设计基于注意力机制的卷积准循环神经网络模型，充分地捕获和增强对剩余时间预测任务有决定性影响的时间和空间特征信息，以提高业务流程中上下文事件之间的关联性，从而识别不同事件对业务流程剩余时间预测任务的重要程度。在BPIC_2012_A、BPIC_2012_O、BPIC_2012_W等事件日志数据集上的实验验证了该方法的有效性和可行性，结果表明，相比传统剩余时间预测方法，该方法的预测准确度平均提高约20%，有助于提升业务流程剩余时间的预测质量。

关键词: 业务流程, 剩余时间预测, 深度学习, 轨迹聚类, 卷积准循环神经网络

Abstract: The existing remaining time prediction methods only focus on the time feature information that plays a decisive role in the remaining time prediction task, and do not consider the impact of spatial feature information and heterogeneous event logs on the prediction task, resulting in a decrease in prediction accuracy.This study proposes remaining time prediction method based on trajectory clustering.The similarity between different trajectories is regarded as a distance measure, and clustering is performed to cluster trajectories with different lengths in the event log to reduce the complexity of the event log and to refine the structure.For the remaining time prediction task of business process, Convolutional Neural Network(CNN) and Quasi-Recurrent Neural Network(QRNN) are combined, and bidirectional and attention mechanisms are introduced to design a Convolutional Quasi-Recurrent Neural Network(CQRNN) model based on the attention mechanism to fully capture and enhance the time and spatial feature information, which significantly affect the remaining time prediction task.The design improves the correlation between contextual events in business process, so as to identify the importance of different events to the business process time prediction task.In experiments labeled as BPIC_2012_A, BPIC_2012_O, and BPIC_2012_W, event log datasets demonstrate the effectiveness and feasibility of the proposed method.The results show that the prediction accuracy of this method is improved by about 20% on average compared with the traditional remaining time prediction method, which is helpful to improve the prediction quality of the business process remaining time.

Key words: business process, remaining time prediction, deep learning, trajectory clustering, Convolutional Quasi-Recurrent Neural Network(CQRNN)

中图分类号: 

• TP311.5