Recursive object-centric process discovery with fitness and rediscovery guarantees

Abstract

Digitized processes are an integral part of businesses to achieve their operational goals. Process mining aims to analyze such processes based on digital track records. These so-called event logs can be extracted from commonly used information systems. Traditionally, process mining remains limited as it requires individual business objects to define isolated process executions. However, in real-life processes, it is much more common to observe multiple business objects of various types that interact with each other. Object-centric process mining acknowledges this fact and fundamentally enhances traditional process mining techniques to analyze processes with interacting business objects. This analysis often requires a process model that describes the interacting control flows of all objects. Discovery algorithms are used to automatically construct such models from logs. For further steps, it is important that the discovery algorithm rediscovers a process model that is equivalent to the underlying business process. If it is impossible to rediscover the full behavior of a process, the discovered model should at least be fitting for the observed behavior. However, no existing object-centric discovery technique guarantees rediscovery or fitness. In this paper, we introduce a new algorithm that recursively constructs tree-structured object-centric models. We prove that our approach produces fitting models for all logs and rediscovers a subclass of processes within the used representational bias from logs that are sufficiently complete and free of incorrect recordings, unobsered or duplicated activities and loops. We evaluate our approach on a range of public logs and find it to produce fitting models in feasible runtime.

The authors' abstract, as published at the source. Information Systems, 2026 · DOI ↗