Abstract: Remaining time prediction is a core task of business process predictive monitoring.Existing predictive models are usually built from offline,so it usually takes a long time to build or update the models which tend to degrade with streaming events.For this reason,a method of online prediction model construction and remaining time prediction was proposed.Based on hybrid transition systems that incorporate multiple abstract mechanisms,an incremental prediction model construction algorithm was proposed to ensure the model up to date.The prediction confidence was defined to represent the qualities of the prediction results,and a confidence-based prediction algorithm was proposed.The prediction algorithm expanded the candidate set of prediction values by looking back the history to improve the prediction accuracy.The volatility of prediction was defined to measure the temporal consistence of prediction.Experiments on multiple public data sets had been conducted,and the results showed that the proposed method outperformed the existing methods in terms of precision and volatility.

Key words: remaining time prediction, volatility, process mining, real-time data

摘要： 鉴于现有剩余时间预测模型所采用的离线构建方式周期长、更新速度慢,在流式事件分析中容易老化,提出一种在线预测模型的构建及剩余时间预测方法。融合多种抽象机制得到复合变迁系统作为预测模型,设计了增量式预测模型学习算法,以保证模型的实时性;基于统计理论定义预测信度,给出基于预测信度的剩余时间预测算法,采用轨迹回顾机制增强模型的预测能力;定义波动性衡量预测结果在时间维度上变化的幅度;通过在多个公开数据集上与已有方法对比,表明所提方法在预测准确性和波动性上均具有明显的优势。

关键词: 剩余时间预测, 波动性, 过程挖掘, 实时数据