A Web-based Framework for the Evaluation of Predictive Process Monitoring Techniques
Name
Stanislav Mõškovski
Abstract
Business process management (BPM) focuses on optimizations of various activities within the organization, with respect to key performance indicators (KPI). An important task among BPM-related activities is process monitoring which aims to make sure that business processes comply with KPIs.
Process monitoring can be performed either offline, using historical data to analyze process execution in the past or online, i.e. analyzing event streams in real-time to identify the problems as soon as they arise. Predictive monitoring is an emerging type of online process monitoring that uses historical data to construct a predictive model using various machine learning methods and then applies this model to a live event stream in order to predict the future performance of ongoing process cases. Various techniques have been proposed to address typical predictive monitoring problems, such as whether this ongoing case will finish on time or what activity will be executed next in the case.
Even though many of these techniques have publicly available software implementations, they typically target one specific predictive monitoring problem. Furthermore, due to variations in evaluation procedures (different data splits, different evaluation metrics reported, etc.), users do not have a readily available way to compare predictive accuracy across multiple techniques.
Finally, such solutions are targeting experienced users and also consume a lot of users hardware resources to run the simulations. In this thesis, we have built a web application that allows users with various degrees of expertise in the subject to train, validate and compare models to predict multiple KPIs, using a wide range of predictive monitoring techniques proposed in related work.
Moreover, the models can be exported for further use. This application runs all of the computations on the server side, thus eliminating the need for the powerful hardware to construct the models. We compare our solution with existing implementations and highlight clear distinctions and differences.
Process monitoring can be performed either offline, using historical data to analyze process execution in the past or online, i.e. analyzing event streams in real-time to identify the problems as soon as they arise. Predictive monitoring is an emerging type of online process monitoring that uses historical data to construct a predictive model using various machine learning methods and then applies this model to a live event stream in order to predict the future performance of ongoing process cases. Various techniques have been proposed to address typical predictive monitoring problems, such as whether this ongoing case will finish on time or what activity will be executed next in the case.
Even though many of these techniques have publicly available software implementations, they typically target one specific predictive monitoring problem. Furthermore, due to variations in evaluation procedures (different data splits, different evaluation metrics reported, etc.), users do not have a readily available way to compare predictive accuracy across multiple techniques.
Finally, such solutions are targeting experienced users and also consume a lot of users hardware resources to run the simulations. In this thesis, we have built a web application that allows users with various degrees of expertise in the subject to train, validate and compare models to predict multiple KPIs, using a wide range of predictive monitoring techniques proposed in related work.
Moreover, the models can be exported for further use. This application runs all of the computations on the server side, thus eliminating the need for the powerful hardware to construct the models. We compare our solution with existing implementations and highlight clear distinctions and differences.
Graduation Thesis language
English
Graduation Thesis type
Bachelor - Computer Science
Supervisor(s)
Ilya Verenich
Defence year
2018