CAPTR: A Context-Aware Personalized Travel Place Recommendation System Using Heterogeneous Graph Transformers

Abstract

Travel recommendation is hard in destination-rich countries such as Sri Lanka. Whether a traveler is satisfied depends on fixed preferences, but also on the weather, the budget, and the many relationships in tourism data. Standard collaborative and content-based recommenders flatten every user-location interaction into one homogeneous relation. They mostly ignore context, so they personalize poorly, stumble on new users and destinations, and react badly to real conditions. This paper describes CAPTR, a context-aware travel place recommender. It represents users, destinations, and their relation types as a heterogeneous graph, then learns node embeddings with a Heterogeneous Graph Transformer. An adaptive gate weighs content and collaborative signals for each request, and a Bayesian optimization step tunes the per-feature importance weights rather than by hand. A re-ranking layer reads a domain-expert weather-activity compatibility matrix and live forecasts to adjust results as conditions change, while a budget filter built on scraped accommodation prices drops what the user cannot afford. New users and destinations join the trained graph through feature-similarity neighbors, so the model serves them without retraining. On 776 travelers and 72 Sri Lankan destinations it reaches an NDCG@10 of 0.694 and a Hit@10 of 1.0, beating all eight baselines under one cold-start protocol.