<p>Cross-domain sentiment analysis aims to address the problem of insufficient labeled data by transferring invariant knowledge across domains. Existing studies focus on unimodal domain transfer, but modal differences hinder the transfer of domain information and limit the acquisition of domain-invariant knowledge in multimodal data. Therefore, we propose a domain-adaptive transfer network (DATN) for multimodal cross-domain sentiment analysis. The joint representation is acquired through a bidirectional visual–textual interactive fusion network, and adversarial discriminative domain adaptation is employed to learn the marginal domain shared knowledge in the joint representation. The performance of multimodal domain adaptive modules aligns with conditional distributions. Extensive experiments on public and self-constructed datasets demonstrate the effectiveness of the model and show that self-constructed datasets have the potential to serve as a new benchmark. Compared with the best-performing method, the model’s accuracy on public datasets increased by 3.6% and 8.2%, respectively, and on self-built datasets by 9.2% and 3.3%, respectively.</p>