Large Language Models for simplifying directly-follows graphs in support of self-regulated learning

Abstract

Directly-Follows Graphs (DFGs) derived from educational process mining provide rich evidence of students' interaction patterns in virtual learning environments, but their complexity often hinders interpretation by instructors and researchers. This study investigates how Large Language Models (LLMs) can support learning analytics by transforming complex DFGs into simplified models and interpretable summaries of student behaviour, with a specific focus on self-regulated learning (SRL). We analyse the actual event log of a compulsory undergraduate course in a Computer Science degree (87 students, ~140k events), discovering a DFG using PM4Py, and compare several GPT-5 variants and reasoning configurations for the task of simplifying this model. The LLM is prompted to remove less relevant transitions while preserving core behaviour. We examine the structural impact of simplification and qualitatively analyse behavioural patterns in a representative simplified DFG. Results reveal a spectrum of trade-offs: smaller models tend to over-prune, whereas larger ones produce more conservative reductions with higher coverage. Despite substantial simplification, the selected GPT-5 configuration preserves navigation, assessment and grade-monitoring patterns that can be interpreted in terms of planning, performance and reflection phases of SRL. We argue that such human-readable process summaries can help instructors detect less effective behaviours, redesign course workflows and scaffold SRL, positioning LLM-assisted DFG simplification as a promising bridge between algorithmic process mining and educational decision-making.