Recommendation systems are widely used everywhere and have a great influence on daily life. Aiming to train an ideal recommendation system, a massive of use-item interactive pairs should be provided. However, the dataset obtained is usually sparse, which might result in an over-fitting model and be hard to obtain the satisfying performance. In order to address this problem, the cross-domain recommendation is raised. Most of the existing methods on cross-domain recommendation systems borrow the ideas of the conventional unsupervised domain adaptation, which employ the feature alignment or adversarial training methods. Hence they can transfer the domain-invariant interests of users from the source to the target domains, e. g., from the Douban Movies to the Douban Books. However, since the network structures vary with different recommendation platforms, the existing methods on cross-domain recommendation systems straight forwardly extract the domain-invariant representation may entangle the structure information, which may result in the false alignment phenomenon. Furthermore, the previous efforts ignore the noise information behind the graph data, which further degenerate the experimental performance. In order to address the aforementioned problems, this paper brought the causal generation process of graph data into the cross-domain recommendation systems, it used the semantic latent variables of each node to calculate the relationships between users and items via disentangling the semantic latent variables, domain latent variables and noise latent variables. Experiments show that the proposed method yields state-of-the-art performance on several cross-domain recommendation system benchmark datasets.