The paper "Incomplete Graph Learning via Attribute-Structure Decoupled Variational Auto-Encoder" addresses the challenge of incomplete graph data, where node attributes are often missing. Traditional Graph Neural Networks (GNNs) struggle with incomplete graphs due to their assumption of fully observable node attributes. The authors propose ASD-VAE (Attribute-Structure Decoupled Variational Auto-Encoder), a neural model that integrates attribute and structural representations into a shared latent space. This approach allows for robust handling of high missing rates by leveraging the shared latent space to impute missing attributes. The model 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 decoupled into separate views, and the reconstruction loss of each view is calculated. Finally, missing attribute values are imputed from this learned latent space. Extensive experiments on four real-world datasets demonstrate the superior performance of ASD-VAE compared to state-of-the-art methods in label prediction and attribute estimation tasks. The model also shows improved resilience against skewed and biased distributions caused by missing information, making it more effective for downstream graph machine-learning tasks.The paper "Incomplete Graph Learning via Attribute-Structure Decoupled Variational Auto-Encoder" addresses the challenge of incomplete graph data, where node attributes are often missing. Traditional Graph Neural Networks (GNNs) struggle with incomplete graphs due to their assumption of fully observable node attributes. The authors propose ASD-VAE (Attribute-Structure Decoupled Variational Auto-Encoder), a neural model that integrates attribute and structural representations into a shared latent space. This approach allows for robust handling of high missing rates by leveraging the shared latent space to impute missing attributes. The model 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 decoupled into separate views, and the reconstruction loss of each view is calculated. Finally, missing attribute values are imputed from this learned latent space. Extensive experiments on four real-world datasets demonstrate the superior performance of ASD-VAE compared to state-of-the-art methods in label prediction and attribute estimation tasks. The model also shows improved resilience against skewed and biased distributions caused by missing information, making it more effective for downstream graph machine-learning tasks.