Trace Encoding Techniques for Multi‐Perspective Process Mining: A Comparative Study

Abstract

Process mining (PM) comprises a variety of methods for discovering information about processes from their execution logs. Some of them, such as trace clustering, trace classification, and anomalous trace detection require a preliminary preprocessing step in which the raw data is encoded into a numerical feature space. To this end, encoding techniques are used to generate vectorial representations of process traces. Most of the PM literature provides trace encoding techniques that look at the control flow, that is, only encode the sequence of activities that characterize a process trace disregarding other process data that is fundamental for effectively describing the process behavior. To fill this gap, in this article we show 19 trace encoding methods that work in a multi‐perspective manner, that is, by embedding events and trace attributes in addition to activity names into the vectorial representations of process traces. We also provide an extensive experimental study where these techniques are applied to real‐life datasets and compared to each other.