.png){kind=icon}
.png){kind=icon}
Return
MVGAE-AC: Multi-view graph autoencoder for attribute completion
.png){kind=icon}
PRE
AI
DOI：10.1016/j.neucom.2025.131753.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Heterogeneous graphs, composed of multiple types of nodes and edges, have been widely adopted in various domains due to their rich semantic information. However, the sensitive nature of data and the incompleteness of data collection often result in a substantial amount of missing node attributes, which severely degrade the performance of downstream tasks. Existing attribute completion methods have partially mitigated the impact of missing attributes on downstream tasks. However, the existing heterogeneous graph neural networks (HGNNs) with attribute completion still face two challenges: (1) Most attribute completion models rely on a pre-training process, which adversely affects downstream task performance. (2) Existing models primarily focus on learning local feature information, while neglecting the influence of global structural information. Furthermore, current models fail to fully leverage the embeddings obtained from the encoder, which restricts expressiveness of the generated attributes. To address these issues, we propose a novel end-to-end attribute completion network, named multi-view graph autoencoder for attribute completion. Specifically, we first mask a portion of node attributes. Secondly, an encoder is utilized to obtain multi-view embeddings from both local feature view and global structural view. Then, the graph structure is reconstructed through a structure decoder and a re-masking strategy is employed to enhance the expressiveness of the reconstructed attributes. Finally, the completed attributes serve as input to a HGNN model for downstream tasks. Extensive experiments conducted on three real-world datasets demonstrate that our proposed method consistently outperforms existing state-of-the-art methods across multiple evaluation metrics. The source code is available at: https://github.com/mbjiii/MVGAE-AC .
Journal
IF:
6.5
Papers:
2.5W
Citations:
6.5W