Abstract: Evolution is inevitable in the business process from the perspective of time dimension.The purpose of drift detection in process mining is to find out the points of changing time in the running log,and then divide different business processes before and after evolution.However,the current drift detection algorithms need to specify the evolution features that should be monitored manually during implementation,which has the problems of low accuracy and long time consuming as well as increases unnecessary burden for users.A detection algorithm based on Graph Convolutional Network was proposed.The traces were transformed into an activity graph,and all features of the traces were characterized by the node information and topological structure of the graph.The global characteristics of the active graph were obtained by using the aggregation characteristics of graph convolution.All information of the activity graph was expressed by adding a virtual node.Finally,Euclidean distance and K-nearest neighbor algorithm were used to detect the position of drift points.The validity of the proposed method was proved by the experimental tests on real public datasets.

Key words: concept drift, business evolution, graph convolutional network, process mining

摘要： 从时间维度来看,演化是业务过程的必然性。流程挖掘中漂移检测的宗旨是找出运行日志中的变化时间点,进而划分出演化前后的不同业务过程。然而,现有漂移检测算法在执行过程中,都需要人为指定需要监控的演化特征,给用户增加了使用负担。同时,还存在准确率低和耗时较长问题。为此,提出一种基于图卷积网络的检测算法:首先,将日志轨迹转为活动图,利用图的节点信息和拓扑结构来刻画日志所有特征;接着,利用图卷积的聚合特性获取到活动图的全局特征;然后,通过增加虚拟节点来表达活动图的所有信息;最后,采用欧氏距离和k-近邻算法来检测漂移点位置。通过真实公开数据集实验测试,表明了本文方法的有效性。

关键词: 概念漂移, 业务演化, 图卷积网络, 流程挖掘