With the acceleration of digital transformation in the hotel industry, traditional static models based on collaborative filtering and content recommendation are difficult to meet the needs of users’ real-time behavioral preference changes, and have problems such as low cold start efficiency, poor adaptability to dynamic scenarios, and lack of multi-objective optimization. This study proposes a reinforcement learning recommendation framework based on deep Q network (DQN), which constructs a multi-dimensional state space of users and hotels, designs a dynamic reward function, adopts a dual network architecture and experience replay mechanism, and realizes strategy optimization based on online incremental learning. Tests on a hotel data set containing 870,000 users and 4.2 million interaction records show that compared with traditional methods, this model has increased the 7-day retention rate of new users to 63.2%, the peak click-through rate of real-time recommendations has reached 41.6%, and the average order conversion rate has remained at 56.2%. Experiments have shown that this model significantly improves the effectiveness of personalized recommendations in dynamic scenarios through the sequential decision-making mechanism of reinforcement learning, providing a feasible solution for the optimization of hotel intelligent service systems.