Multi-source unsupervised domain adaptation (MUDA) is a challenging research problem, which aims to leverage the knowledge from multiple labeled source domains to adapt to an unlabeled target domain for better inference performance. Despite many methods having made efforts to extract common domain-invariant representations across all domains to improve adaptation ability, diverse cross-domain distribution discrepancies and imprecise decision boundaries have not been well addressed. To deal with such challenges, in this work, we propose a Domain Graph-Structured Multi-Source Domain Adaptation with Dual Integration named DGSDA. Specifically, the proposed method contains three core modules: (1) Input integration, where the combined source domain is integrated by the multi-source domains to reduce the distribution discrepancies; (2) Domain graph structure module, constructed through feature extraction and domain graph embedding, focusing on aligning domain-specific distributions to learn multiple domain-invariant representations for optimal category decision boundaries; (3) Decision integration, aiming to assign different weights to multiple classification results to efficiently explore useful knowledge. Extensive experiments on real-world image classification tasks demonstrate that the proposed DGSDA achieves expressive domain adaptation performance.