This paper proposes ASD-VAE, a neural model for imputing missing attributes in graphs. The model decouples attribute and structure representations into a shared latent space, enabling robust tolerance to high missing rates. ASD-VAE first separately encodes attributes and structures, then learns a shared latent space by maximizing the likelihood of the joint distribution of different view representations. The shared latent space is then decoupled into separate views, and the reconstruction loss of each view is calculated. Missing values are imputed from this learned latent space, enhancing resilience against skewed and biased distributions. The model is evaluated on four real-world incomplete graph datasets, demonstrating superior performance in label prediction and attribute estimation tasks. ASD-VAE outperforms existing methods, particularly in scenarios with high missing rates. The model also introduces a Katz-GCN component to improve attribute estimation and downstream tasks. The results show that ASD-VAE achieves high-quality attribute imputation and effective node classification, even under high missing rates. The model's ability to handle missing data and its effectiveness in downstream tasks make it a promising approach for graph learning.This paper proposes ASD-VAE, a neural model for imputing missing attributes in graphs. The model decouples attribute and structure representations into a shared latent space, enabling robust tolerance to high missing rates. ASD-VAE first separately encodes attributes and structures, then learns a shared latent space by maximizing the likelihood of the joint distribution of different view representations. The shared latent space is then decoupled into separate views, and the reconstruction loss of each view is calculated. Missing values are imputed from this learned latent space, enhancing resilience against skewed and biased distributions. The model is evaluated on four real-world incomplete graph datasets, demonstrating superior performance in label prediction and attribute estimation tasks. ASD-VAE outperforms existing methods, particularly in scenarios with high missing rates. The model also introduces a Katz-GCN component to improve attribute estimation and downstream tasks. The results show that ASD-VAE achieves high-quality attribute imputation and effective node classification, even under high missing rates. The model's ability to handle missing data and its effectiveness in downstream tasks make it a promising approach for graph learning.