Goal-Driven Process Variant Categorization: Using LLMs to Bridge Behavior and Business Objectives

The Variant Categorization Challenge in Process Mining

Process Discovery rarely produces a single, coherent process structure in practice. Real-world business systems generate a large number of process variants — execution paths that differ in subtle ways across instances of the same business process. To make these variants analyzable, the standard approach is to first cluster variants by structural similarity, then assign business meaning to each resulting group.

The problem is that these clusters are not derived from the organization's business objectives — they are mechanically induced from the structural characteristics of process data. This means analysts must manually interpret each group and consolidate variants into business-meaningful categories. As the number and complexity of variants grow, this labor-intensive step becomes increasingly unwieldy.

Process mining is fundamentally a technique for automatically extracting business process models from event logs. An event log consists of a series of timestamped activity events recorded by a system, where each complete execution record is called a case. When different cases follow different sequences of activities, process variants emerge. In real enterprise systems (such as ERP or CRM platforms), even a single business process can produce dozens or even hundreds of variants due to factors like customer type, exception handling, or individual operator habits. Traditional structural similarity clustering typically relies on edit distance or topological similarity of control-flow graphs to merge variants — measures that capture only the surface form of execution paths, not the business meaning behind activities. This is the fundamental bottleneck of purely data-driven classification: two structurally similar variants may correspond to entirely different business scenarios, while structurally dissimilar variants may align closely in terms of business objectives.

Inverting the Workflow: Define Objectives First, Categorize Later

The core idea proposed in this arXiv paper is to invert the traditional categorization workflow. Rather than "cluster first, assign meaning later," the approach advocates "define objectives first, categorize variants later."

The method unfolds in several steps: researchers first construct a goal model for the organization that predefines the categorization axes — meaning the dimensions and criteria for classification are determined upfront by business objectives, rather than passively discovered from data. Each process variant is then transformed into a textual narrative that describes its behavioral characteristics in natural language.

The LLM as a Semantic Bridge

The critical step is handed to a large language model: the LLM interprets each variant's textual narrative within the context of the goal model and assigns it to the most appropriate category. The paper emphasizes that it is precisely the LLM's semantic reasoning capability that connects low-level process behavior to the high-level business objectives defined by analysts.

The elegance of this design is that it sidesteps the limitations of purely structural similarity clustering. Two structurally similar variants may not serve the same business objective; conversely, structurally dissimilar variants may belong to the same class in terms of business semantics. The LLM's involvement allows categorization to genuinely revolve around "business intent" rather than "data shape."

Goal models are well-established tools in requirements engineering and business analysis, with common forms including the i* framework and the KAOS method. Their core idea is to express organizational intent hierarchically as goals, sub-goals, and their dependencies — rather than directly describing system behavior. In this paper's context, the goal model serves to predefine "which business dimensions we care about" — for example, "did the case go through exception approval?" or "does the processing path satisfy compliance requirements?" — thereby anchoring the categorization axes in business semantics rather than leaving them to be determined by data characteristics alone. The step of converting process variants into textual narratives is equally critical: this transformation maps structured activity sequences into natural language, enabling the LLM to leverage semantic knowledge accumulated during pretraining to understand process behavior, without requiring special adaptation to process graph structures.

Experimental Evaluation and Findings

The research team instantiated this approach end-to-end and evaluated it on three publicly available event log datasets. These datasets differ significantly in scale and behavioral diversity, providing a reasonably robust test of the method's adaptability.

The evaluation yielded several noteworthy findings:

• Partitions guided by the goal model differ noticeably from those produced by unguided induction. This demonstrates that the goal-driven approach genuinely changes categorization outcomes, rather than merely replicating structural clustering in a different form.
• The method responds to controlled modifications of declared alternatives in the goal model. In other words, when analysts adjust category definitions within the goal model, the categorization results change accordingly — reflecting the method's controllability and interpretability.
• The trade-off is the additional upfront effort required to author a goal model. This is the prerequisite cost that this approach must bear compared to automated clustering.

Value and Limitations

From a value perspective, this work represents a meaningful direction: embedding LLM semantic understanding into specific steps of process analysis, enabling machines to take on the interpretive work that previously demanded extensive manual judgment from analysts. When the number of variants is large and manual consolidation becomes impractical, this kind of automated semantic categorization has genuine practical appeal.

At the same time, its limitations deserve clear-eyed acknowledgment. Constructing a high-quality goal model requires domain expertise and human effort; whether this upfront cost is worthwhile depends on the scale of variants and frequency of analysis in a given context. Furthermore, since the paper positions the LLM as the core of semantic reasoning, the stability, reproducibility, and performance of classification results across goal models from different domains still require broader validation.

For readers interested in the intersection of process mining, Business Process Management (BPM), and LLM applications, this research offers a clear paradigm example: use a predefined goal model to constrain categorization dimensions, and use a large language model to complete the mapping from behavior to semantics — thereby aligning process variant analysis more closely with the business objectives organizations actually care about.