The Research On Remaining Processing Time Prediction Of Business Process Based On Multi-Head Attention LSTM Adversarial Network

With the application of predictive process monitoring in different fields,such as economy,digital health,business process management,and IT infrastructure monitoring,it has increasingly become one of the main research hotspots in the field of process mining.As an application of predictive analysis,process prediction mainly uses models learned from historical event logs to predict the evolution of ongoing process cases,such as the activity suffix of the case or the remaining processing time.Recently,several predictive process monitoring methods based on deep learning have been proposed to solve the problem of low accuracy of activity suffix prediction or remaining processing time.However,due to insufficient training data or poor models,these methods cannot solve the current problems well.The main focus of this article is the activity suffix and the remaining time prediction.Their tasks are to obtain the most likely follow-up activity and the remaining time before the end of the case based on the current activity.Aiming at the problem of insufficient mining of semantic information and spatial structure information between activities in the existing deep learning models for business process prediction,and low prediction accuracy in predicting activity suffixes and remaining processing time,this paper proposes a based on Multi-head Attention LSTM Adversarial Network model.The main research work of this paper is as follows:(1)Aiming at the problem that the existing suffix and remaining time prediction models are insufficient for the mining of semantic information and spatial structure information between activities,this paper proposes an encoding-decoding architecture based on Multi-head Attention LSTM,which deeply extracts the features of time series data and better capture dependencies between events across time.(2)Aiming at the problem of low prediction accuracy of the existing suffix and remaining time prediction models,this paper proposes a based on Multi-head Attention LSTM Adversarial Network model,which will confront the generator and the discriminator,so as to obtain predictions that are indistinguishable from the real situation.In turn,the accuracy of predicting the activity suffix and remaining processing time is improved.The worst-case prediction accuracy of this method is at least the same as that in a non-adversarial environment.(3)In order to verify the accuracy of the prediction of the multi-head attention LSTM confrontation network model,this article conducts experiments on four public data sets,and uses the Damerau-Levenstein distance and Mean Absolute Error(MAE)to predict the activity suffix and the remaining processing time respectively.The results are evaluated and compared with the baseline.The experimental results show that the proposed method improves the prediction accuracy compared with the baseline model,and the method is more robust because it is not affected by the fluctuation of the prefix length and is more applicable.