To deal with the diversity of feature space in multiple heterogeneous domains, we describe a fusion multi-domain semantic topics and syntax classes model to extract latent long-range topics and short-range classes in multi-domain. The model fuses the information from multiple heterogeneous domains. There are two layers in the fusion model. Firstly, in every sub-domain, the model uses the hidden Markov model with latent dirichlet allocation (HMM-LDA) to extract sub-domain topic and class features. Secondly, the fusion model combines the multiple sub-domain models to extract the whole heterogeneous domain features. The experiments use the international universal recommendation datasets. The results show that our method outperforms three typical recommendation algorithms, and reduces the Mean Absolute Error by an average of 0.313 and 0.182 on MovieLens and Book-Crossing dataset separately. According to the method that extracting single domain features, our proposed method reduces the Mean Absolute Error by an average of 0.065 and 0.099 on MovieLens and Book-Crossing dataset separately. The result also exhibits that our method outperforms other method in the sparse data.