Unsupervised Task Recognition from User Interaction Streams

Abstract

User interaction events can give an accurate picture of tasks executed in a process, since they capture work performed across applications in a detailed manner. However, such data is too low level to be used for process analysis directly, since the underlying tasks are typically not apparent from individual events. Therefore, several task-recognition techniques were recently proposed that are able to abstract user interaction data to a higher level. However, these techniques work in an offline manner, requiring user interaction data to be stored in event logs. Such storage is often infeasible, though, due to the data’s sheer volume and its privacy-sensitive nature. While this can be avoided by analyzing user interaction data in a streaming manner, existing task-recognition techniques cannot be applied to such settings, since they require multiple, post-hoc passes over the entire data collection. To overcome this, we propose the first approach for unsupervised task recognition from user interaction streams. For a given stream, our approach continuously identifies task instances, groups them according to their type, and emits task-level events to an output stream. Our evaluation demonstrates our approach’s efficacy and shows that it outperforms two baseline approaches.