Current recommendation systems for tourism struggle to capture the dynamic changes in user preferences. Therefore, this study proposes a knowledge graph embedding technique that combines dynamic mapping matrices to construct a tourism recommendation model. Meanwhile, bidirectional long short-term memory networks and node-level attention mechanisms are introduced to enhance the modeling ability for dynamic changes in user behavior. The experimental results on the YAGO11k dataset showed that the accuracy of the training set reached 99.4%, and the model had excellent training performance and generalization ability. In the evaluation indicators of the knowledge graph, the average ranking and average reciprocal ranking were 452 and 0.430, significantly better than the baseline, with a hit rate of up to 29.1%. This model provides an effective solution for personalized travel recommendations.