Proactive Business Process Outcome Risk Monitoring Based On Machine Learning

With the development of information technology,enterprise organizations monitor processes of production and operation,and record business event logs through business process information management systems.Knowledge information of business processes can be obtained with process mining technology by analyzing event logs,so that enterprises can improve and adapt business process execution outcomes to reduce various costs.During business process execution,predictive business process monitoring methods predict business process execution activities and future execution status using machine learning commonly.When a specific process event is executed,the monitoring method needs to proactively judge whether there is a risk of failure in the future execution outcome according to the current prediction,so as to determine whether the monitoring system triggers an alarm.This alarm notifies the process manager in time to adapt and improve the current business process.In order to reserve operational space for adaptation and improvement,and to ensure high prediction accuracy of the monitoring method,the trigger control process of proactive business process adaptation needs to discover process outcome failure as early and accurately as possible.If the monitoring method fails to predict process outcome failure,it may lead to huge losses for the enterprise;but misprediction of the outcome failure will introduce unnecessary adaptation and increase the process burden.With the continuous execution of process events,the accuracy of prediction for the monitoring method will be enhanced with more event information.But it will gradually reduce the operational space for process adaptation.Therefore,proactive process adaptation needs to balance the earliness and accuracy when predictive monitoring.In order to meet the requirements of proactive monitoring and process cost control,this paper uses a variety of machine learning techniques to propose a process outcome risk monitoring method,mainly based on ensemble gated graph neural networks and proximal policy optimization.Firstly,by combining the control structure information of the process execution and the context data integrated into the process event attributes,using gated graph neural networks model,this article proposes a modeling prediction method which converts sequential recorded process traces into process graph structures and constructs process activities and their transfer relationships as nodes and directed edges for learning.The prediction method improves information representation and information capture capability,and contributes to the prediction accuracy requirements.Secondly,based on the interpretability requirements in process monitoring,information is extracted from the node vectors learned by the graph neural networks model to summarize the graph-level prediction results,and parameters are extracted in the prediction output layer to calculate the impact of each node on the prediction results as the relevance scores of the process activities.Using relevance scores with the combination of prediction results and process execution information can achieve explainable analysis of business processes.Finally,this paper uses ensemble learning based on prediction model to calculate the reliability estimates of the prediction results,and combines process execution information with activity relevance information to build reinforcement learning environments for using proximal policy optimization algorithm.The process adaptation trigger method based on reinforcement learning is proposed to improve the adaptive control triggering ability with the process state and reward information modeling in proactive process adaptation.With connecting the process prediction model,the algorithm architecture for process outcome risk monitoring is formed.This paper conducts experiments by using event logs that record real business processes.The evaluation selects appropriate prediction accuracy metrics and process cost function,analyzes the impact of activity relevance in prediction results in comparative experiments,and verifies the effectiveness of the prediction method in terms of accuracy and the performance of adaptation trigger method in process cost control.