An intelligent analysis and personalized recommendation system for tourist preferences based on multi-source data

A rising number of tourist recommendation systems (TRS) are tailoring their suggestions to each user by mining their location-based social networks and user-generated behavioral data. However, existing approaches often fail to integrate diverse tourist data sources efficiently, account for changing tastes over time, and provide recommendations with supporting explanations. A hybrid graph-transformer-based model, the Multi-Source Tourist Preference Recommendation (MSTPR) can learn multimodal tourism data to grasp complicated spatial-social relationships; this work presents it as a solution to these inadequacies. The architecture jointly represents long-term trip preferences and short-term contextual behaviors via parametric gating, a temporal preference modeling mechanism, a hybrid graph-transformer module, and dynamic context-sensitive embeddings. To further investigate how social, environmental, and contextual factors affect recommendation outcomes via attention weighting, a multi-head attention mechanism is employed. To assess the model’s performance, it is tested on a real-world multimodal tourist dataset that includes user interactions with POIs, contextual travel information, and atmospheric conditions. A combination of experimental assessment using different ranking and classification metrics and a temporal train-validation-test split of 70/10/20 is utilized. According to the results, MSTPR consistently outperforms baseline recommendations and models. It achieves an AUC of approximately 0.999 and a 7.5% improvement in accuracy, along with higher performance in precision, recall, F1-score, NDCG, MAP, Hit Rate, Coverage, and Intra-List Diversity. Another quantitative piece of information provided by the attention-based interpretability mechanism is the relative importance of social, temporal, and spatial elements influencing tourists’ preferences. Personalized tourist recommendations that are both accurate and easy to understand can be generated using the proposed hybrid graph-transformer structure, as shown by these outcomes.