This paper takes the behavioral trajectory dynamics of tourists in tourist attractions as the research object on the premise of big data and adopts mean filtering technology to preprocess tourists’ behavioral trajectories. After that, LSTM and RNN are used to analyze the preferences of long-term and short-term tourists and explore the spatio-temporal factors affecting tourists’ behavioral trajectories. Finally, vector embedding and hierarchical attention mechanisms are applied to recommend intelligent services to tourists for points of interest. The results show that the culling of influencing factors reduces the model’s recommendation performance and affects the tourists’ decision to visit the points of interest. The MALS model has the best recommendation effect at TOP = 10. In this paper, tourists are clustered into three categories: category one (52%): spending, cognition, and education are on the lower end of the scale, family trips are the main focus, and food is extremely preferred. Category 2 (21%): higher spending, cognition, and education, mostly traveling with friends or alone, preferring humanities and history, entertainment activities, catering and food, and intelligent scenic services. The third group (27%): mainly undergraduates aged 18-25, mostly traveling with friends or couples, with lower expenses, preferring tour guide services and natural landscapes.