Extracting Process Instances from User Interaction Logs

Abstract

Thoroughly documenting digital business processes in a company is a crucial and necessary, yet cumbersome task. However, having detailed documentation of one's processes in a modelling language like Business Process Model and Notation (BPMN) can prove very useful regarding process optimization or automation, employee training and on- and offboarding. Process and task mining frameworks try to ease the creation of process documentation by automatically generating it based on transaction or user interaction data with the system. These approaches often have the disadvantage of not covering the whole process due to a variety of possible execution paths and their habit of not continuously recording process data. We propose an extension to the task mining tool Desktop Activity Mining (DAM) which allows to capture data continuously over several hours and therefore not miss any important cases that might not occur very often. This approach also limits the influence of human errors when recording process data with certain frameworks for documentation purposes and provide the possibility of an improved degree of automation. We evaluate the approach on real-world data to show its feasibility and application in practice. We used a combination of already existing algorithms and created our own. By classifying 332 unique user interactions, we end up with 76 different equivalence classes. Evaluating the algorithm, we achieved a classification correctness of 70% in two datasets.