基于注意力的日志属性特征权重分配模型研究  

Research on Attention Mechanism-based Event Log Attribute Feature Weighting Model

作　　者：步卓伦 巢翌[1] 李晓龙 BU Zhuo-lun;CHAO Yi;LI Xiao-long(Beijing Institute of Control and Electronic Technology,Beijing 100038,China)

机构地区：[1]北京控制与电子技术研究所,北京100038

出　　处：《计算机仿真》2025年第2期389-394,共6页Computer Simulation

摘　　要：为了解决大型业务系统中用户操作产生的事件日志间聚合度低与相关性弱的问题,针对聚类使用的数据集属性特征进行优化,提出了一种基于注意力机制的事件日志属性特征权重分配模型(Attribute Feature Weighting Model,AFWM),通过分析不同事件日志中属性特征的重要程度来提高事件日志分类后数据集的聚合度与关联度,为后续流程挖掘的轨迹聚类提供基础。设计了一种新颖的特征权重分配方法,方法采用Transformer架构来获取事件日志属性特征间的相互依赖关系,根据上下文关系为属性特征分配权重值,以此区分不同特征间的重要程度,提高事件日志聚类准确度。通过大量公开现实事件日志集对上述方法进行评估,最终证明所提的方法可以对复杂业务系统产生的事件日志集进行正确的聚类。To address the issue of low aggregation and weak correlation among event logs generated by user operations in large business systems,this paper optimizes the dataset attributes used in clustering.We propose an Attribute Feature Weighting Model(AFWM)based on the attention mechanism,which enhances the aggregation and association of the dataset after event log classification by analyzing the importance of attribute features in different event logs,providing a foundation for subsequent process mining trajectory clustering.A novel feature weight allocation method is designed in this study,utilizing the Transformer architecture to capture the interdependencies among event log attributes.Weights are assigned to attribute features based on contextual relationships,distinguishing the importance of different features and improving the accuracy of event log clustering.We evaluated our method using a large number of publicly available real-world event log datasets,ultimately proving that our approach can correctly cluster event logs generated by complex business systems.

关 键 词：流程挖掘 注意力机制 日志聚类 特征挖掘 

分 类 号：TP391.9[自动化与计算机技术—计算机应用技术]