Locally optimized inductive process discovery

Abstract Process mining provides techniques to analyse business processes based on digital execution track records, called event logs. A key step in that analysis is the construction of a process model that describes the workflow of the process. Process discovery algorithms automatically obtain these models from event logs. For any subsequent analysis it is important that the discovered model fits the behaviour of the log precisely, without overfitting the input data or using behaviorally unsound constructs. Achieving this goal under real-life circumstances with potentially incomplete and partially incorrect input logs remains challenging. State-of-the-art algorithms for this purpose typically rely on removing large parts of the process from the final model or utilize domain knowledge in the form of suitable input parameter choices. In this paper, we investigate if these two shortcomings of state-of-the-art process discovery techniques can be addressed without compromising on soundness guarantees and model quality. For this purpose, we introduce a new process discovery algorithm, called OptIMIIst, which extends the Inductive Miner framework for the recursive construction of sound process models with integer linear programs to optimize local mining decisions. Our evaluation demonstrates that OptIMIIst produces models with competitive quality profiles in comparison to other approaches with soundness guarantees without relying on excessive filtering or suitable parameter inputs.