電腦科學與資訊工程科 Computer Science & Information Engineering
190037 Saudi Arabia
Scout: A Personality-Aware LLM-Based Travel Recommendation System Tailored to Saudi Arabia
Saudi Arabia tourism sector is undergoing rapid expansion, evolving into a new paradigm of innovation and cultural exchange, this transformation is propelled by Saudi Vision 2030. Despite this growth, current tourism mainstream systems do not fully capitalize on this rapid expansion or showcase its unique offerings, providing travel itineraries that are fundamentally generic, non-personalized, inflexible, and incapable of capturing individual traveler preferences. Thereby undermining both the global visibility of Saudi tourism and its potential to deliver culturally rich, personalized experiences. To address this real-world problem, this research introduces Scout, as a next generation LLM-powered travel assistant that can form comprehensive, end-to-end travel itineraries to Saudi Arabia that is systematically tailored to the user’s individual psychobehavioral traits, contextual preferences, and real-time data. The system operates through a multi-layered architecture in which natural language interaction initiates a concurrent processing pipeline involving a core generative LLM, dedicated psychobehavioral analysis models, and real-time retrieval modules. Outputs from these heterogeneous components are dynamically integrated into Scout's novel mechanism NeuroFusion Layer (NFL). It's an active integration mechanism modeled after how human biological neurons in the brain connect and pass information along synaptic path. It serves as a pre-layer in the base transformer LLM architecture and fuses these outputs to a fused normalized vector represented as z_norm ∈ R^(d_p). This representation is injected into the LLM embedding space as an additive bias, allowing the model to incorporate behavioral, contextual, and retrieved knowledge directly into the generation process. Compared to a vanilla transformer baseline, the NeuroFusion Layer (NFL) integrated to Scout's architecture reduced training loss and perplexity by over 60%, demonstrating more stable learning and improved contextual modeling. This architectural enhancement further reduced hallucinations and constraint violations during itinerary generation by 32% , while achieving a 92% alignment between generated travel plans and users’ psychobehavioral profiles, validating robust and consistent personalization. Thus, Scout novel architecture has a significant impact on culturally informed, personalized tourism experiences in Saudi Arabia, supporting the goals of Saudi Vision 2030.