Cancer’s biological heterogeneity poses significant challenges for subtyping and personalized treatment. Multi-omics data, including gene expression, DNA methylation, and miRNA expression, offers potential for identifying distinct cancer subtypes but is often high-dimensional and incomplete. Conventional methods relying on imputation or sample exclusion risk losing critical biological information [10]. This paper proposes a robust framework using Multiple Similarity Network Embedding (MSNE) to integrate incomplete multi-omics data without imputation. By constructing similarity networks with Gaussian kernels, normalizing local neighborhoods, and fusing them into a unified graph, patient embeddings are generated via random walks and clustered using unsupervised algorithms. Evaluated on a dataset with 30% simulated missingness, the approach achieves high-quality clusters (Silhouette=0.595, NMI=0.710 for k = 3), demonstrating robustness and advancing precision oncology.