Abstract: The next event prediction task is one of the research focuses of predictive process monitoring,and the existing deep learning-based prediction methods suffer from long training time,large amount of parameters and high hardware requirements to meet the dynamic nature of business processes.To address these problems,a Sampling-based Next Event Prediction (SNEP) method based on log sampling was proposed.Specifically,the importance of traces was measured by calculating event importance and direct-following activity relationship importance,and some important traces were extracted to represent the original event log.The prefixes of trace were recoded using the One-hot coding approach and a three-layer Long Short Term Memory(LSTM) network prediction model applicable to the next event prediction task was designed.Experiments were conducted in six real event logs to investigate the effectiveness of the proposed method and the effect of different sampling rates on the prediction results of the model.The results showed that the pre-sampled next event prediction method had improved prediction accuracy and efficiency in each event log,which could help practitioners to achieve next event prediction tasks efficiently.

Key words: business process, prediction of next event, event log sampling, deep learning, long short term memory network

摘要： 下一事件预测任务是预测性流程监控的研究重点之一,针对现有基于深度学习的预测方法存在训练时间过长、参数量过大、对硬件要求过高等无法满足业务流程动态性的问题,提出一种基于日志采样的下一事件预测方法(SNEP)。通过计算事件重要性和直接跟随活动关系重要性来衡量轨迹重要性,抽取部分重要轨迹表示原事件日志;采用One-hot编码方式对轨迹前缀重新编码,并设计了适用下一事件预测任务的三层长短期记忆网络（LSTM）预测模型。在6个真实事件日志中进行实验,探究所提方法的有效性和不同采样率对模型预测结果的影响,结果表明预先采样的下一事件预测方法在各事件日志中的预测准确率和效率均有提升,验证了该方法的优越性。

关键词: 业务流程, 下一事件预测, 事件日志采样, 深度学习, 长短期记忆网络