<p>Personalized travel itinerary recommendation aims to suggest ordered sequences of points of interest (POIs) that align with user preferences while respecting temporal and spatial constraints. Although traditional methods have made progress in next-POI prediction and route planning, they often rely on static user preferences and overly simplistic spatial assumptions, overlooking contextual variability and non-linear distance effects. Additionally, many approaches decouple POI selection from itinerary construction and depend heavily on limited behavioral signals, failing to capture richer feedback such as ratings, dwell times, or click histories. To address these limitations, we propose an intention-aware deep learning framework that integrates diverse user signals into a unified itinerary planning model. A novel fusion-aware encoder assimilates both explicit and implicit user feedback to uncover latent preferences driving POI choices. The model combines a multi-head self-attention transformer to capture the sequential and temporal dynamics of user behavior, with a graph convolutional network (GCN) that models complex co-visitation patterns among POIs. These components are jointly trained to select optimal POIs and construct itineraries that respect user-specific constraints and time budgets. Our custom itinerary optimization heuristic balances user interest with travel and waiting costs, directly aligning recommendation with practical route planning. Extensive experiments on multiple real-world datasets - including theme parks and urban tourism - demonstrate that our model consistently outperforms state-of-the-art baselines across a range of metrics. It produces more coherent and temporally balanced itineraries, effectively reducing queue times and improving user satisfaction. Our results highlight the value of multi-signal fusion and contextual modeling in capturing nuanced user intent, setting a new benchmark for personalized itinerary generation.</p>