Weight-Driven Nearest Neighbor Algorithms for Effective Data Linking

Personalized recommender systems are transforming the travel industry by delivering tailored experiences based on user preferences. However, new user recommendations remain challenging due to the cold start problem, where limited historical data restricts personalization accuracy. To address this, we introduce an AI-driven, adaptive recommendation framework that combines Large Language Models (LLMs) with Reinforcement Learning (RL) to refine travel suggestions dynamically based on evolving user interactions. Our system leverages PhoBERT-based natural language processing (NLP) for Vietnamese text classification, enabling deep semantic understanding of user queries. We represent each user through a dynamic feature vector that captures historical behaviors, contextual factors (weather, events, seasonal trends), and real-time interactions. When a new user engages with the system, a top-K influence model identifies relevant attributes from existing users to generate context-aware recommendations. Furthermore, by integrating LLMs with RL, our system continuously learns from user feedback, refining its suggestions over time. This hybrid approach bridges historical behavior modeling with adaptive real-time learning, significantly improving personalization accuracy. Experimental results demonstrate an 91% satisfaction rate, highlighting the effectiveness of influence-based user modeling and reinforcement learning in mitigating the cold start problem. Our findings advance AI-powered recommender systems, offering a scalable solution for emerging travel markets, with potential applications beyond Vietnam.