Noise-resistant discovery of identifier-sound object-centric process models with dynamic identity relations

Abstract Business objects involved in real-life processes are often tracked by information systems in digital execution records, called object-centric event logs. Object-centric process mining encompasses techniques to analyze processes based on such logs. This analysis relies on the discovery of a process model that reflects the control flow of all object types and the interactions between them. In real-life processes, these interactions are often dependent on the object identities. In an order-to-cash process, for example, there is an identity-based link between the sales order, the delivery and the corresponding invoice. Existing object-centric modeling formalisms either ignore these identity relations, treat them as statically known in advance, or capture them without guaranteeing important behavioral soundness properties. In this paper, we formalize ten types of dynamic identity relations in object-centric event logs, selected based on a corresponding literature survey. Then, we extend the modeling formalisms of object-centric Petri nets and object-centric process trees to support these types of identity relations. Additionally, we introduce a noise-resistant algorithm to automatically discover corresponding process models from object-centric event logs. We prove that our extension preserves important soundness properties and, under the absence of noise, strictly improves model quality criteria. We evaluate our approach on a range of logs and observe the consistent discovery of dynamic identity relations in feasible runtime. Additionally, we find our discovery technique to detect dynamic identity relations proportional to the selected noise parameter threshold.