|M.Sc Student||Berkenstadt Guy|
|Subject||Performance Inference in Queueing Networks with Partial|
|Department||Department of Industrial Engineering and Management||Supervisor||PROF. Avigdor Gal|
Measuring key performance indicators, such as queue lengths and waiting times, using event logs serve for improvement of resource-driven business processes with emphasis on the queue aspect, when the resource is scarce. However, existing queue mining techniques assume the availability of complete life cycle information, including the time a case was scheduled for execution (aka arrival times) in addition to the start and completion of execution. Yet, in practice, the arrival times may be missing for a large portion of the recorded cases for various reasons.
In this thesis, we propose a methodology to address missing life-cycle data by incorporating predicted information in business processes performance analysis. Our approach builds upon techniques from queueing theory, which produce an estimation for queue length distribution. We offer a design for a feature space extracting from this distribution and additional information from the event log with missing data. Then, we leverage supervised learning to accurately predict performance indicators based on our design for a feature space.
Our experimental results using both synthetic and real-world data demonstrate the effectiveness of our approach. This effectiveness is especially evident in the real-world dataset that does not comply with the queueing theory assumptions. But, it is most likely to achieve this effectiveness from the additional features in the event log.