Abstract: Information systems generate a substantial amount of business logs during the execution of business processes,which often encompass the processes of multiple departments.If process discovery is directly applied to raw logs generated by such information systems,it can result in an excessively complex and unreadable process model.Therefore,it is necessary to pre-cluster the logs into multiple sub-logs,ensuring that related processes within the same department are consolidated into individual sub-logs.Trace clustering is an effective strategy to mitigate the issue of overly complex processes.A novel method for clustering event logs utilizing graph convolutional networks was proposed,which comprised the following three phases:①the K-means algorithm was used to divide the raw log into a large number of sub-logs,and each sub-log would be transformed into a directed graph;②the graph convolutional networks were adopted to capture the features in each graph,then a metric matrix was used to record the Euclidean distance between any two directed graphs;③the hierarchical clustering method was used to merge these graphs until the number of clusters met the user's needs.The method had been validated on five public datasets,and extensive experiments demonstrated the superiority over other existing methods.

Key words: process discovery, graph convolutional networks, event logs, hierarchical clustering, directed graph

摘要： 信息系统在执行业务过程时会产生大量业务日志,这些日志中包含了多个部门的业务流程。如果直接对信息系统产生的日志进行流程发现,会得到一个十分复杂且不可读的流程模型。因此,需要预先对日志数据进行聚类得到多个子日志,使得同一个部门的相关流程尽量在一个子日志中。轨迹聚类是一种能够缓解流程过于复杂问题的有效策略。本文提出了一种利用图卷积网络对事件日志进行聚类的新方法,该方法包含以下3步:①使用K-means算法将原始日志划分为大量子日志,并将每个子日志转换为有向图;②采用图卷积网络来捕捉每个图的特征,然后使用度量矩阵记录任意两个有向图之间的欧氏距离;③利用层次聚类方法合并这些有向图,直到簇的数量满足用户的需求。通过大量实验在5个公开的数据集上进行了验证,表明本文方法优于现有方法。

关键词: 流程发现, 图卷积, 事件日志, 层次聚类, 有向图