Optimization of Personalized Tourism Services via Multimodal Mobile Interaction

Abstract

In mobile tourism scenarios, the inherent contradiction between the demand for real-time personalized services and the constraints of terminal computing power, battery life, and network fluctuations presents a significant challenge. Existing solutions commonly suffer from fragmented interaction entry points, delayed contextual responses, and excessive reliance on cloud computing. To address these issues, a three-tier collaborative optimization framework, encompassing on-device perception, edge coordination, and dynamic orchestration, was proposed. Within this framework, modality-level attention pooling was employed to achieve adaptively weighted fusion of visual, auditory, and sensory signals. A physical-digital fused mobile context tensor was constructed, and temporal encoding was utilized to capture the evolutionary patterns of user intent. Service scheduling was formulated as a constrained Markov decision process, and a resource-aware deep reinforcement learning algorithm was applied to derive an optimal edge-device collaborative service orchestration policy. To ensure deployability on mid-range mobile devices, complementary mechanisms of dynamic network pruning and pseudo-label distillation were incorporated. Experimental results on a realworld scenic area dataset demonstrated that the proposed method achieved a Top-1 intent recognition accuracy that was 9.7% higher than the next-best baseline. The end-to-end average response latency was reduced by 30.3% compared to reinforcement learning-based orchestration schemes of a similar type. Furthermore, the energy consumption per inference was decreased by 42% relative to the original model, with an accuracy degradation controlled within 3%. This work provides a viable technical paradigm for optimizing smart tourism services in mobile environments.