Context-Based Activity Label-Splitting.

The information systems used in companies store event data describing the historical execution of the processes they support. Process mining covers the automated analysis of such data, generating insights that may ultimately lead to process improvement. A core branch of process mining is process discovery, dealing with event-data-based automated discovery of process models. In practice, the same activity may often be executed in a significantly different context, e.g., in a vaccination program, multiple vaccine doses are typically provided at different points in time. Process discovery algorithms assume that all executions of the same activity are to be mapped onto the same modeling element. Consequently, the presence of repeated activity executions under different contexts typically leads to underfitting discovered process models. To this end, activity label-splitting algorithms have been proposed to relabel the recordings of the same activity occurring in significantly different execution contexts. Yet, the state-of-the-art label-splitting algorithm adopts a trace-level-mapping strategy, yielding inferior results in the presence of loop constructs and infeasible computation time. Therefore, this paper proposes a novel label-splitting preprocessing technique that overcomes these issues. Our experiments confirm that our newly proposed label-splitting algorithm outperforms the state-of-the-art.