Semantically Informed Structural Process Model Similarity with Graph Neural Networks

Organizations often maintain multiple process models that describe similar or related business processes. Assessing the similarity between these models is important for applications such as process comparison, harmonization, standardization, and model retrieval.

Existing approaches to process model similarity typically consider different dimensions, including semantic, structural, and behavioral similarity. However, process models that represent very similar processes may look structurally quite different. The same functionality may, for example, be represented by a single activity in one model and by an entire process fragment in another. Conversely, models with similar structures do not necessarily represent similar processes. This makes it difficult to determine which structural differences are actually relevant when assessing process model similarity.
This thesis investigates semantically informed structural process model similarity using Graph Neural Networks (GNNs). The central idea is to start from known correspondences between process elements or fragments and abstract away their original activity labels. Corresponding elements can, for example, be represented through shared abstract identifiers, allowing the analysis to focus on how corresponding parts are structurally organized rather than on their textual labels.

Based on these abstracted process models, you will investigate whether GNNs can learn representations that capture meaningful process model similarity despite differences in structure and modeling granularity. In particular, the thesis explores whether a learned graph-based similarity measure can distinguish structural differences that arise from alternative ways of modeling the same or corresponding process fragments from differences that indicate genuinely different process structures.

What You Will Do

• Review the literature on process model similarity, with a particular focus on structural similarity, graph-based representations, and applications of Graph Neural Networks.
• Design a graph-based similarity approach. Develop a suitable representation of process models that incorporates known correspondences between process elements and fragments and investigate how GNNs can be used to derive model similarity.
• Implement the approach as a prototype that can automatically calculate similarity between process models using one or more GNN architectures.
• Evaluate the approach against existing similarity measures, with particular attention to process models that differ in structure and modeling granularity.
• Investigate the role of structural representation. Analyze how different representations of process models and known correspondences affect the learned similarity measure.
• Analyze successes and failures to understand which types of structural differences can be captured by the GNN-based approach and where limitations remain.
• Write up the findings, discussing the implications of semantically informed structural similarity for process model comparison, harmonization, and standardization.

Your Profile

• Genuine interest in business process management, process modelling, and applications of machine learning
• Ability and willingness to implement and experimentally evaluate a GNN-based prototype
• Experience with Python or willingness to become proficient with it
• Interest in working with Graph Neural Networks, graph representations, and structured data
• Prior exposure to BPMN, process modelling, graph algorithms, or machine learning

The topic can be scoped for a Bachelor's or a Master's thesis. A Bachelor's thesis may focus on implementing and evaluating a selected GNN architecture and process model representation, while a Master's thesis can investigate alternative graph representations, learning strategies, and similarity formulations in greater depth. The exact research questions and scope will be aligned during the proposal process before registration.

Please get in touch with a short email including your motivation for writing a thesis at our chair (in the body of your email), your CV, a current transcript of records, and the planned start/finish date: rehse@ifi.uzh.ch.