LLMs Abstract User Actions into Interpretable Workflows Across Diverse Applications.

Understanding how users interact with digital applications is crucial for product improvement, but raw interaction logs are often too granular and noisy to yield clear insights. Traditional deep learning methods for clustering user actions into higher-level activities have struggled with noise sensitivity and cross-application generalization. A new framework called WorkflowView addresses these limitations by leveraging large language models (LLMs) to abstract these low-level action sequences into meaningful, high-level activities. This method has been tested across three distinct and challenging sequential tasks, proving its effectiveness and versatility. For instance, it can reconstruct task descriptions from browser logs with high semantic similarity, predict student dropout from MOOC logs with high accuracy using minimal examples, and perform privacy-preserving analysis of AI tool integration in document workflows. This research highlights LLM-based abstraction as a robust and efficient way to convert behavioral data into actionable insights, with practical considerations for deployment efficiency and user privacy.