Extracting Multi-Destination Tourism Flows Using LLM-Driven Process Mining: A Case Study of Structural Shifts in Thai Domestic Travel via Social Media Group Before-, During-, and After COVID-19

Traditional tourism tracking methods, particularly those reliant on GPS, often face limitations regarding data intrusiveness and restricted spatial scope.  Addressing this, this study introduces a novel data pipeline integrating Large Language Models with process mining to analyze unstructured social media data. The framework is validated through a case study of Thai domestic tourism, mapping trajectory shifts across pre-, during-, and post-pandemic phases. Analysis of 2,577 itineraries reveals a distinct structural evolution: from 2019’s linear, mono-centric flows to 2021’s rigid "safety loops" where hotels served as isolated sanctuaries. By 2023, a complex, hybrid mesh emerged, establishing hotels as permanent logistical hubs while religious tourism remained the sole consistent anchor. These findings demonstrate the pipeline's capability to uncover intricate behavioral transitions effectively.

Consequently, this research contributes a scalable, nonintrusive methodology for monitoring nationwide multidestination trajectories. To facilitate reproducibility, the complete dataset, pipeline code, and data flow graph results are published on GitHub1.