GUIDE: Generative User Intent Dual-Encoder for tourism recommendation

Tourism and point-of-interest (POI) recommendation is inherently intent-dependent and context-sensitive, yet many existing methods primarily rely on user–item interactions or short-term behavioral sequences, making it difficult to capture the rich semantics of natural-language travel demands. To address this limitation, we propose GUIDE, a G enerative U ser I ntent D ual- E ncoder framework. We first distinguish a user’s current intent from long-term preference , and construct target-free semantic user contexts from pre-target histories and contextual metadata to avoid label leakage. GUIDE then uses an open-source large language model strictly as a semantic encoder to represent user-side intent and item-side POI content in a shared embedding space. On top of this backbone, a bidirectional contrastive alignment module improves fine-grained user–item matching, and a variational latent preference module models the uncertainty and multi-faceted nature of travel interests through latent preference sampling. This design combines efficient approximate nearest-neighbor retrieval with diverse and robust re-ranking. We evaluate GUIDE primarily on a genuine tourism/POI check-in dataset, Foursquare-TKY, and further examine its cross-domain behavior on Yelp and MovieLens, comparing it against popularity-based, collaborative, graph-based, sequential, self-supervised, POI-specific, and LLM-enhanced baselines. Under a statistically validated protocol with five random seeds, GUIDE consistently improves top- K ranking on the POI dataset, with significant gains over the strongest baseline (paired Wilcoxon signed-rank test, p < 0.01 ), while auxiliary results on Yelp and MovieLens suggest that the semantic dual-encoder design also transfers to broader recommendation settings. Ablation, parameter-matched, efficiency, fairness, and case analyses further verify the contribution of each component, and a dedicated analysis maps each gain onto the prior finding it confirms, extends, or qualifies. Our conclusions are limited to the evaluated public datasets and are not claimed to generalize to all users or destinations. Code and preprocessing scripts are released for reproducibility.