Abstract
Predictive Process Monitoring is a growing branch of Process Mining that focuses on the analysis of historical execution traces to provide the user with predictions about the unrolling of a partially executed process instance. Typical examples of predictions of the unrolling of a process instance relate to its completion time, to the fulfillment or violation of a certain condition, or to the sequence of its future activities. While Predictive Process Monitoring is a comparatively new area of research, its state-of-the-art techniques already provide the fundamental machinery able to deliver reliable and highperformance predictions, where performance is usually measured in terms of standard notions such as accuracy and time. At the beginning of my thesis the available techniques focused on building reliable and accurate techniques, without focusing on the actual usefulness of these techniques.
In this thesis, we tackle the aspect of improving the usefulness of Predictive Process Monitoring prediction techniques. Usefulness is a broad term and many different things can be done to make Predictive Process Monitoring more useful. We instantiate it to achieve the following three goals:
(i) We increase usefulness by increasing the accuracy of an already trained predictive model. We do this in two ways. First, we explore the use of explanations to enhance the predictive model accuracy. Second, we study how to keep a predictive model up-to-date when new data becomes available.
(ii) We increase usefulness by focusing on predictions’ explanations. We do this by evaluating how understandable and useful are the explanation plots for the business analysts in the Predictive Process Monitoring scenario. (iii) We increase usefulness by facilitating access to Predictive Process Monitoring techniques. We do this by defining an architecture capable of supporting the usage of Predictive Process Monitoring capabilities and we implement the defined architecture in an opensource tool. 7 To the best of our knowledge, this work represents the first take on usefulness of Predictive Process Monitoring among the dimensions of (i) an increased accuracy by exploiting model explanation techniques and update; (ii) the understandability of the explanation of predictions; and finally
(iii) the availability of a Predictive Process Monitoring tool. In detail, this thesis proposes a first approach for increasing the predictive model accuracy using explainability; a first study on how to keep a model up-to-date when new data becomes available over time; a first user evaluation of how understandable and useful are the explanation plots for the business analysts in Predictive Process Monitoring; and a definition of an architecture capable to provide all of the above in one cohesive opensource tool.