Research On Process Concept Drift Detection And Algorithm For Decision Point Mining

The significant improvement in computer's data storage,data transmis-sion,and processing capabilities has considerably increased the number of event data in process-aware information systems,such as workflow management sys-tems,business process management systems.How to effectively employ thesedata as the treasure to mining useful information becomes a research area for tremendous researchers for nearly a decade,which helps to establish the subj ect of process mining.Although a great number of research outcomes have been applied in industry,there are still lots of challenges in the subject of process mining.The process mining manifesto[1]169,written by a group of more than 75 people involving more than 50 organizations,lists 11 important challenges in process mining,including finding,merging,and cleaning event data,dealing with concept drift,combing process mining with other types of analysis and cross-organizational mining.This paper focuses on dealing with concept drift and decision mining.Considering dealing with concept drift,multiple researchers have devel-oped meaningful results,however,they only concentrate on offline detecting using statistical hypothesis testing and sliding windows.Because of the fea-ture of simultaneously handling a large amount of event data and high latency,those approaches could not use in online scenarios.To solve this problem,we present a framework for online process concept drift detection,which consists of three model libraries to store the current model,the candidate models and the previous models.The current model is depicted as a benchmark of the current process traces which are analyzed for detecting concept drift.The candidate models are models that are lately detected.The previous models are models that have been replaced and can help for detecting the type of the concept drift.According to the framework,we introduce an approach for online concept drift detection which uses the methods of relation extractions and the concept of pro-cess model precision.The experiments show that our framework can provide real-time process concept drift detection with high precision.Considering the research on decision mining,existing approaches are di-vided into three steps.Firstly,they use already existed process mining algo-rithms to discover control-flow process models.Secondly,according to the control-flow process models,they specify the decision points.Finally,using machine learning technologies,such as decision mining algorithm,the informa-tion about decisions has been identified.However,process mining algorithms my encounter the problems of poor fitness,precision,generalization,or simplic-ity.These certainly influence the accuracy of decision mining.In this paper,we use the relationship between activities to determine the decision points,which highly improve the precision of the algorithm and reduce the time complexity.