Research And Implementation Of Interpretable Prediction Method For Business Process Activities

Many business processes are generated during the operation of the enterprise organization,and the execution process of these business processes is recorded in business logs.Through the monitoring of these business logs,potential problems during process execution can be predicted in advance,and these problems can be actively managed and resolved,thereby improving the timely response capability during process execution,which is called predictive business process monitoring.Business process activity prediction is an important research field in predictive business process monitoring.The goal is to predict the most likely next activity attribute of a running process based on the completed historical event log.With the rapid development of deep learning technology,more and more research applies deep learning technology to the field of business process activity prediction,and these technologies show excellent performance.However,in recent years,the most advanced business process activity prediction technologies have complex internal structures.While these technologies can provide high-quality predictions,it is difficult to explore the reasons behind their predictions.This also results in a lack of user trust in such technologies and hinders their practical application in sensitive fields.Based on the above background,the research on the interpretability of business process prediction technology is of great significance.This paper mainly studies and implements the interpretability model in the scenario of business process activity prediction,evaluates the interpretability of the model.And provides corresponding solutions for process adaptations based on interpretable models.The main work is as follows:(1)For the interpretability problem of business process activity prediction,this paper proposes an interpretable business process activity prediction model based on LSTM and attention mechanism.The hidden state of different attributes in events is learned by exploring the structure of LSTM recurrent neural network,and then the attention mechanism is used to simulate the generation process of the next activity.Attribute importance and temporal importance are derived as the interpretability for predictions.This paper proposes a perturbation-based interpretable evaluation method to evaluate interpretable information.The experimental results show that the method in this paper can achieve higher activity prediction accuracy,and the interpretable information provided corresponds to the real situation of the business process.(2)For the application of interpretable model in process adaptations,this paper proposes a triggering process adaptations method based on ensemble interpretable prediction and reinforcement learning.This method combines the two machine learning paradigms of reliability estimate based on ensemble interpretable models and triggering process adaptations based on reinforcement learning to form an adaptive triggering process adaptations scheme that adds interpretable information.Experimental results show that the method proposed in this paper can bring higher accuracy and lower process adaptations costs.(3)On the basis of the method proposed in this paper,this paper designs and implements a business process interpretable prediction prototype system.The prototype system can predict the next activity of the ongoing trace and provide interpretable information about the prediction results.At the same time,the prototype system can judge whether the current process needs to be actively adapted,so that users can take timely intervention decisions.