Abstracting Cross-Domain Action Sequences into Interpretable Workflows

The article introduces WorkflowView, a framework leveraging large language models (LLMs) to abstract low-level action sequences from digital application usage into high-level activities. It demonstrates effectiveness across three tasks: zero-shot task description reconstruction from browser logs (semantic similarity of 0.91), few-shot dropout prediction from MOOC logs (weighted F1 score of 0.90 with five examples), and privacy-preserving analysis of AI tool integration in Microsoft Word. This approach offers a robust method for transforming noisy interaction data into interpretable insights, which is crucial for practitioners aiming to enhance digital products based on real user behavior.