|Ph.D Student||Senderovich Arik|
|Subject||Queue Mining: Service Perspectives in Process Mining|
|Department||Department of Industrial Engineering and Management||Supervisors||Professor Avigdor Gal|
|Professor Avishai Mandelbaum|
|Full Thesis text|
Business processes are supported by information systems that record process-related events in event logs. Process mining is a maturing research field that aims at discovering useful information about the business process from these event logs. Process mining can be viewed as the link that connects process analysis fields (e.g. business process management and operations research) to data analysis fields (e.g. machine learning and data mining).
This thesis relates to process mining techniques that aim at answering operational questions such as `does the executed process as observed in the event log correspond to what was planned?', `how long will it take for a running case to finish?' and `how should resource capacity or staffing levels change to improve the process with respect to some cost criteria?
Prior to this thesis, process mining techniques overlooked dependencies between cases when answering such operational questions. For example, state-of-the-art methods for predicting remaining times of running cases considered only historical data of the case itself, while the interactions among cases (e.g. through queueing for shared resources) were neglected. The independence assumption is plausible in processes where capacity always exceeds demand. However, in service processes, which are prevalent in numerous application domains (e.g., healthcare, banking, transportation), multiple customer-resource interactions occur, and customers often compete over scarce resources. Consequently, the central argument of this thesis is that for service-oriented processes, process mining solutions must consider case interactions when answering operational questions.
The main contribution of this research thesis is the start of bridging a noticeable gap between process mining and queueing theory. To this end, we introduce queue mining (a term coined in this thesis), which is a set of data-driven methods (models and algorithms) for queueing analysis of business processes.
We demonstrate that considering the queueing perspective yields improved solutions to operational problems and enables solutions to problems that were not addressed earlier in the literature. Our queue mining techniques address the problems of prediction (delays and total times in the process), conformance to schedule (planned vs. actual), and process improvement (via production policy optimization). We demonstrate the effectiveness of these techniques with experiments on real-world data that comes from three domains: banking (a bank's call center), transportation (city buses), and healthcare (an outpatient hospital).