Abstract: Process prediction monitoring can predict the subsequent steps and related attributes of business instance to prevent the risk in future and take timely interventions.As one of prediction tasks,process remaining time prediction can avoid the timeout risk.However,with growth over time and business scale,the business execution process will change accordingly.This requires that the prediction model can be continuously updated to capture these changes.In addition,enough input information is needed to distinguish the difference before and after the change,and the prediction model should have sufficient fitting and generalization ability.To tackle above challenges,we introduce a framework to support incremental logs for remaining time prediction.Specifically,we first proposed a feature self-selection strategy to provide enough input information,based on selected features constructed the multi-feature prediction model to improve the fitting ability.On that basis,taking the time cycle and the case number as judgment standard of the model update,we present three incremental update mechanism including regular updates,quantitative updates,and comprehensive updates.Based on six real event logs,the incremental update mechanism is simulated on implemented three prediction models.The experimental results verify the effectiveness of the proposed approach and improve the accuracy of the process remaining time prediction.

Key words: process prediction monitoring, remaining time, incremental update, feature selection, multi-feature prediction

摘要： 流程预测性监控通过对业务流程及其属性的预测,预防运行中的实例未来可能会面临的风险,从而及时干预流程。流程剩余时间预测是避免业务超时风险的一项预测任务,然而业务执行是动态的过程,可能会随时间或业务规模的增长而发生变化。这就要求预测模型能够持续更新以捕捉这些变化,同时要有足够的输入信息来区分变化前后的差异,并且预测模型应具有充分的拟合与泛化能力。针对上述问题,本文提出支持增量日志的流程剩余时间预测框架。具体而言,提出特征自选取策略,构建多特征预测模型,丰富预测任务的已知信息,将所得特征组合作为模型输入,提高预测模型的拟合能力。然后,将定期和定量作为模型更新的判断依据,提出定期更新、定量更新和综合更新3种增量更新机制。最后,基于6个真实事件日志,实现了3种不同的预测模型,模拟了增量更新过程。实验结果验证了本文所提方法的有效性,提高了流程剩余时间预测的准确率。

关键词: 预测性监控, 剩余时间, 增量更新, 特征选择, 多特征预测