Investigating the Limits of Graph Foundation Model in Real-World Travel Recommendation Systems

PAKDD 2025 - GLFM Workshop

Abstract

Graph foundation models (GFMs) have demonstrated remarkable potential in capturing intricate relational patterns, achieving state-of-the-art results in numerous graph-centric tasks. However, their real-world applicability remains underexplored in highly domain-specific contexts, such as travel recommendation. In this paper, we present a comprehensive evaluation of GFMs for large-scale travel recommendation tasks using a bipartite user–destination dataset of 86,761 travelers within South Korea. We compare representative GFM against both conventional graph-based methods and vector-based methods. Contrary to the prevailing expectation that GFMs should outperform traditional architectures, our empirical findings reveal that domain-specific constraints can dilute the benefits of extensive multi-hop message passing, leading to suboptimal performance. Our work highlights a critical need to validate GFMs against domain-specific constraints, offering a roadmap for their future adaptation and optimization in real-world applications.

Nayoung Lee
Nayoung Lee
M.S. Student

First MT in my life happened here!

Donghun Lee
Donghun Lee
Assistant Professor

Connecting artificial intelligence and mathematics, in both directions.