Object-centric process mining (OCPM)

What is object-centric process mining?

Object-centric process mining (OCPM) is an approach where events are not tied to a single case but to several business objects at once: an order, the related invoice, the delivery and any return. In classical process mining you have to pick one of those objects as the case up front. With OCPM you do not: an event can be linked to an order plus two invoices plus three deliveries, and all those relationships are kept.

The term was introduced in 2019 by Wil van der Aalst in his paper Object-Centric Process Mining: Dealing with Divergence and Convergence in Event Data (SEFM conference). Van der Aalst leads the Process and Data Science (PADS) group at RWTH Aachen University, the driving force behind OCPM and the OCEL standard.

Why does classical event-log analysis fall short?

A classical event log forces you to pick one case ID per row. As long as your process revolves around that one object, that works fine. But the moment you also bring in invoices, deliveries or payments, you run into two stubborn problems the literature calls convergence and divergence.

Convergence happens when one event belongs to several cases. Think of one invoice that combines three orders. Flatten the log by order number and "invoice created" gets copied three times, so your analysis counts the event three times even though it only happened once.

Divergence is the mirror image. One case contains several instances of the same activity that can no longer be told apart. One order with three deliveries generates three "shipped" events under the same case ID, and the discovered process looks like a loop when in fact there were three parallel sub-processes.

OCPM fixes this at the source, by keeping the multi-object structure in the data model.

How does the object model work?

An object-centric event log has three building blocks, rather than one row per event with one case ID.

Objects are the business entities that move through your process: orders, order lines, invoices, deliveries, parcels. Each object has a type and its own identifier, with attributes that can change over time.

Events are the activities, each with a timestamp, linked to one or more objects. "Invoice sent" is attached to the invoice and to the orders it settles.

Relationships record how objects and events hang together: event-to-object (this event concerns these objects) and object-to-object (this order contains these order lines). In OCEL 2.0 you can give those relationships a qualifier to describe the kind of link.

The standard that captures this is OCEL (Object-Centric Event Log). OCEL 1.0 was proposed in 2020 and formally published in 2021. OCEL 2.0 appeared in March 2024 and adds changing object attributes and explicit object-to-object relationships. Exchange is in JSON, XML or SQLite.

Example: split shipment with multiple invoices

Imagine you run an e-commerce business. A customer orders three items: a book, shoes and a jacket. The book ships right away, the shoes follow two days later from a second warehouse, the jacket is out of stock and ships only a week later. The customer receives two invoices: one for the book and shoes, one for the jacket. The shoes do not fit and go back, triggering a credit note.

Pick the order number as the case ID and three shipments, two invoices and one return end up jumbled together in the same case (divergence). Pick the invoice number and "order created" gets copied across two cases (convergence). Either choice breaks something.

In an object-centric log you do not have to choose. You model order, order line, delivery, invoice and return as separate object types. "Shipped" happens three times, each time linked to one delivery plus the matching order line. "Invoice sent" hangs off the invoice and the invoiced order lines. The return hangs off the shoes order line, the corresponding delivery and the credit note.

From that single log an OCPM tool draws several perspectives and shows where those cycles meet: where order lines wait for shipment, which invoices are held up by incomplete deliveries, how many returns follow which type of delivery.

When do you need OCPM?

Not every process needs OCPM. If there is one dominant object and a clear case, classical process mining does the job fine. OCPM gets interesting once there are real many-to-many relationships between objects:

• Order-to-cash with partial deliveries or consolidated invoices. One order leads to multiple deliveries, or multiple orders to one invoice. Typical in retail, e-commerce, wholesale and manufacturing.

• Purchase-to-pay with consolidation. One purchase order covers multiple goods receipts on one supplier invoice. Classical process mining pushes you into an argument about which document is the "real" case ID.

• Production processes with batches and serial numbers. A production order splits into sub-batches with numbered products, each with its own quality check and shipment.

• Logistics with containers and parcels. One shipment holds multiple parcels; one parcel bundles items from several orders.

Rule of thumb: if the workshop to choose a case ID turns into half an hour of argument without consensus, that is often a sign the process is object-centric by nature.

Current state of play

OCPM is no longer a futuristic concept, but not yet fully mainstream either. The research comes mostly from the PADS group in Aachen. The field now has a shared standard (OCEL 2.0) and a first generation of tools.

On the open-source side, ProM supports OCEL 1.0 and 2.0 through the OCELStandard plugin, with discovery for object-centric directly-follows graphs (OC-DFG) and object-centric Petri nets. Alongside that, there is the OC-PM web application (by Alessandro Berti) and Python libraries like ocpa and PM4Py with object-centric extensions.

On the commercial side, Celonis is the most visible vendor, with an object-centric data model, prebuilt object and event types for order-to-cash and purchase-to-pay, and connectors for SAP ECC, S/4HANA, Oracle, Salesforce and other source systems. Other process mining vendors follow at varying speed, some in preview, others still case-centric for now. Always check the current vendor documentation before you buy anything off the back of a demo.

For processes with clear many-to-many relationships, OCPM already delivers better analyses than classical process mining today, although a good object-centric log needs more modelling work up front. Start with a process where you know the flat approach is already blocking you.