Social data-driven travel: Integrating NCF and wide & deep learning for personalized experiences (DeepTravelRS)

Abstract

DeepTravelRS is a real-time, large-scale tourism recommender system engineered to personalize travel itineraries by integrating heterogeneous contextual factors—including user profiles, travel modes, seasonality, budgets, and geographic constraints—while sustaining sub-20 ms response latency in production environments. Its hybrid architecture fuses Neural Collaborative Filtering for nonlinear latent user–item interaction modeling, Wide & Deep Learning to capture explicit contextual features and memorization, and a Neo4j knowledge graph encoding semantic and geographic relationships between users, points of interest (POIs), and hotels. Real-time orchestration is achieved through GraphQL serving and Redis caching, delivering a 92 % cache hit rate, while Proximal Policy Optimization dynamically adapts recommendations to evolving travel contexts. Experiments on a TripAdvisor-derived dataset containing 52 000 attractions, 2.1 million graph nodes, and 9 141 cities with temporal and geospatial metadata demonstrate that DeepTravelRS outperforms collaborative filtering and single-model baselines, achieving Precision@10 = 0.82, NDCG@10 = 0.85, and Intra-List Diversity = 0.63. Training on AWS SageMaker with Tesla V100 GPUs reduces computation time from 14 hours to 2 hours, enabling frequent model updates without compromising responsiveness. By uniting latent interaction learning, explicit context modeling, graph-based reasoning, and reinforcement learning within a scalable, production-ready pipeline, DeepTravelRS delivers high-accuracy, diverse, and contextually relevant recommendations, making it well-suited for deployment in large-scale, real-time tourism personalization systems.