Abstract
POWL (Partially Ordered Workflow Language) is a novel process modeling language designed to bridge the gap between the structural guarantees of hierarchical process trees and the flexibility of partial orders. By enabling models to represent complex, non-hierarchical dependencies while remaining "sound by design," POWL facilitates the discovery of process models that are simultaneously simpler and more precise than those generated by traditional methodologies.
Introduction
In Business Process Management (BPM), a persistent trade-off exists between model expressiveness and reliability. Traditional Petri nets, such as Workflow nets (WF-nets), offer high expressiveness but are prone to behavioral issues like deadlocks or unreachable states. Conversely, Process Trees ensure soundness by construction through a hierarchical structure of operators (e.g., choice, loop, sequence), yet they struggle to model non-hierarchical dependencies where activities are concurrent or sequential in ways that do not fit a strict tree structure.
While Partial Orders can describe these complex dependencies compactly, they lack the ability to handle fundamental control-flow structures like choices or loops. POWL was developed to integrate these paradigms, offering a recursive modeling language that maintains soundness while allowing for partial order structures.
Methodology
POWL is defined as a recursive language where a model can be a single activity, a combination of sub-models using control-flow operators (Choice $\times$ or Loop $\circlearrowleft$), or a set of sub-models organized as a partially ordered set (poset). In a POWL poset, edges represent sequential dependencies, while unconnected nodes denote concurrency.
The primary discovery mechanism for POWL is the $IM_P$ algorithm, an extension of the Inductive Miner (IM). $IM_P$ introduces a specific step to "mine for a partial order" when standard tree cuts (sequence, parallel, choice, loop) are not found in the event log. This allows the algorithm to identify local dependencies that traditional tree-based miners would either over-generalize or over-complicate. Importantly, every POWL model can be transformed into a sound WF-net, ensuring compatibility with existing process analysis tools.
Evaluation and Results
The authors evaluated the $IM_P$ algorithm using several real-world event logs, including BPI Challenge logs (2017–2020), Sepsis treatment cases, and Road Traffic Fine Management data. The algorithm was compared against the standard Inductive Miner (IM) across three key metrics:
- Simplicity: $IM_P$ produced simpler models (fewer arcs) in 18 out of 22 test cases by capturing local dependencies more naturally than tree structures.
- Precision: $IM_P$ significantly improved precision. For the BPI-17 log (8 activities), precision increased from 0.37 with IM to 0.68 with $IM_P$, indicating a much tighter fit to the observed behavior.
- Performance: The added complexity of mining partial orders results in a computational trade-off. For example, on the BPI-20-3 log (12 activities), $IM_P$ took 280 seconds compared to less than a second for the standard IM.
Discussion and Future Work
The "soundness by design" property of POWL makes it a robust choice for automated process discovery, avoiding the common pitfall of generating invalid models. By allowing for "partial" concurrency—where some activities have dependencies while others remain concurrent—POWL captures interleaved dependencies that force rigid choices in traditional process trees.
Current limitations include scalability issues due to the brute-force nature of the partial order validation step. Future research aims to optimize this discovery process and explore the integration of Large Language Models (LLMs) to enhance model interpretation and discovery.
Summary updated