Incomplete multi-view multi-label classification, which lies at the intersection of multi-view learning and multi-label classification, has gained considerable attention in recent years due to its broad applicability. However, most existing methods addressing the dual issues of missing views and labels fail to account for the semantic disconnection between labels and feature structures and their incomplete constraint problems. To solve these challenges, we introduce the Dual-Space Structure Preservation with Non-Homogeneous Fusion framework (DSPNF). Unlike conventional approaches, our method jointly models high-order relational structures and semantic correlations within the label space. Specifically, we effectively guide the feature encoding process by employing random walk and graph embedding techniques to impose dual constraints on the topological relationships between instances, thereby significantly enhancing the model’s ability to distinguish high-level semantic features. Furthermore, considering the non-homogeneity in view representation quality, where different views exhibit significant variations in their feature representation contributions to the discrimination task, we propose an uncertainty-driven quality assessment mechanism, which not only precisely quantifies each view’s representation quality but also provides a reliable basis for optimizing cross-view contribution distribution, thereby enhancing classification performance. Our method demonstrates superior performance over leading approaches across five datasets.