Robust drift characterization from event streams of business processes

Abstract

Business processes are prone to change and evolution. Process workers often change the execution of a process in order to adjust to changes in their operational environment, e.g. changes in workload, season or regulations. These process changes are often undocumented and over time may negatively affect process performance. As such, several techniques have been developed for detecting process changes, a.k.a. process drifts, from event logs and event streams, recording the executions of a process. However, detecting a drift without providing explanations on its nature, a.k.a. drift characterization, is not enough to help analysts understand and rectify process performance issues. The existing approaches for drift characterization are limited to changes applied to individual activities. In this paper, we present an automated method for characterizing process fragment changes from event streams. We first adapt a state-of-the-art process discovery technique to work on event stream and use it to discover two process trees from event streams before and after a drift. Each process tree represents a single entry single exit process fragment. As such, we define a set of fragment-based process tree edit operations. We then present two search algorithms for finding a sequence of edit operations with the minimum cost that transforms the pre-drift process tree to the post-drift process tree. At last, we aggregate the identified edit operations as much as possible before reporting them to the user as natural language statements. The hierarchical structure of process trees enables our method to characterize complex changes, such as overlapping changes, and even nested changes. An extensive evaluation on artificial and real-life datasets shows that our method is fast and accurate, and performs significantly better than the state of the art.