Spatio-Temporal Graph Spectral Network for Personalized Itinerary Recommendation

Abstract

The goal of personalized itinerary recommendation is to generate travel routes that closely match each user’s unique preferences and spatiotemporal constraints. This task, however, is complicated by the prevalence of incidental and random “noise” interactions embedded within users’ behavioral histories. Most existing methods process these noisy records directly in the node domain, struggling to reliably separate stable interests from fleeting actions. To address this fundamental challenge, we propose a novel Spatio-Temporal Graph Spectral Network (ST-GSN). Rather than analyzing user behaviors solely in the node space, our approach shifts the perspective into the graph spectral domain. Specifically, we construct for each user a dynamic graph enriched with spatial, temporal, and semantic information, then project their behavioral signals into the spectral domain via the Graph Fourier Transform (GFT). We hypothesize that stable user preferences manifest as low-frequency, energy-concentrated signals, while noise emerges as high-frequency components. Leveraging this property, we design a learnable adaptive filter that precisely isolates and suppresses noise in the spectral space, enabling the extraction of a user’s core intent. The model further incorporates Time2Vec for fine-grained modeling of dwell and travel times, and employs a multi-task learning framework to enhance the robustness of its representations. Extensive experiments on the public Foursquare and Gowalla datasets show that ST-GSN consistently outperforms a suite of strong baselines across all key metrics. Most notably, in the full-corpus ranking scenario that best simulates real-world deployment, the advantage of ST-GSN becomes even more pronounced, demonstrating outstanding performance and resilience in the face of complex, noisy environments.