Graphs are a powerful way of representing data and relationships, with collections of nodes and edges used to describe connections between different entities. In the modern big-data environment, large datasets are often represented as graphs. Deep learning-based graph neural networks (GNNs) provide a powerful approach for modeling such graph-structured data.
GNNs capture relationships in graph data by using a message-passing paradigm, in which the representation of each node is updated by combining information from its neighboring nodes. This allows GNNs to learn low-dimensional representations of nodes that incorporate information from their local graph neighborhoods. Consequently, GNNs have achieved strong performance in a range of graph-related tasks, including social network analysis, link prediction, and graph classification. However, in real-world applications, graph data can be incomplete, with some node features or structural relationships missing. Such incompleteness can degrade the performance of GNNs.
To mitigate this, graph completion learning (GCL) approaches have been proposed that can recover and reconstruct missing node features or structural relationships. Several GCL methods for either feature completion or structure completion have been developed. While effective, existing GCL models heavily rely on a large number of labeled nodes, limiting applications in scenarios where only a small number of nodes are labeled. Moreover, existing GCL methods have largely focused on cases where either node features or graph structure is missing, rather than the more challenging situation in which both are incomplete simultaneously.
To address these challenges, a new study published in Volume 13, Issue 07 of IEEE/CAA Journal of Automatica Sinica on August 03, 2026, presents a more general GCL, termed Extremely Weak Supervision–Robust Graph Completion Network (EWS-RGCN). “ Our approach separates the feature and structure completion into two channels, alleviating the mutual interference between missing node features and structure relationships caused by message passing of GNNs, ” explains author Chengxiang Lei from the Electric Power Research Institute of Guangdong Power Grid Co., Ltd., in China. “ Additionally, we utilize a multi-level contrastive graph mask autoencoder to overcome limitations of limited labels or weak supervision .”
The proposed EWS-RGCN model handles feature and structure completion separately in their respective channels. In the feature channel, missing node features are reconstructed using a trainable parameter matrix that is optimized together with the rest of the model. On the other hand, for structure completion, a personalized PageRank is introduced to reconstruct the missing structure based on existing structural relationships.
“ This bifurcation strategy ensures that channels do not interfere at the initial stage and each channel focuses on extracting pertinent information independently, ” notes Lei.
Next, the model utilizes a multi-level contrastive graph mask autoencoder to extract effective supervision information from the data itself, thus reducing its dependence on labeled nodes. The graph information from the two channels is encoded separately in this scheme. For the feature channel, encoding involves two steps: structure-guided feature diffusion and feature transformation via a multi-layer perceptron (MLP). The former is aimed at effective message propagation during the message-passing process, while the latter handles information transformation. For the structure encoder, positional encoding using graph convolutional networks is employed to generate the structure-channel node embeddings.
For both channels, the decoding process involves masking the embeddings, which are then fed into an MLP decoder. The node embeddings from the two channels are then effectively fused using attention mechanisms to support the final classification. The framework also incorporates an inter-channel information cooperation module to enhance mutual learning between the feature and structure completion channels.
Experiments on six benchmark datasets under different feature and structure missing rates and with limited labeled nodes demonstrated the effectiveness of EWS-RGCN. The approach outperformed existing GCL approaches across all scenarios. “ Our approach can handle both missing features and structures simultaneously, even with extremely limited labeled nodes, ” remarks Lei.
By combining graph completion with extremely weak supervision and separate feature and structure processing, EWS-RGCN provides a potential approach for improving the robustness of GNNs when graph data are incomplete and labeled training data are scarce.
Reference
Title of original paper: Training Robust Graph Completion Networks with Extremely Weak Supervision on Graphs with Incomplete Features and Structure
Journal: IEEE/CAA Journal of Automatica Sinica
DOI: https://doi.org/10.1109/JAS.2025.125906
About IEEE/CAA Journal of Automatica Sinica
A journal of CAA and IEEE, publishing high-quality English-language papers on original theoretical and experimental research and development across all areas of automation. It holds a JCR Impact Factor of 18.3, ranking 1 st (1/88) in the Automation & Control Systems category and placing in the top SCI quartile (Q1). Its CiteScore stands at 27.7, ranking 2 nd in Control and Optimization, in the top 1% for Information Systems, top 1% for Control and Systems Engineering, and top 3% for Artificial Intelligence, with a Q1 quantile overall. On Google Scholar, it has an h5-index of 95, placing it in the top 5 journals in Automation & Control.
Website: https://www.ieee-jas.net/indexen.htm
About Chengxiang Lei from the Electric Power Research Institute of Guangdong Power Grid Co., Ltd
Chengxiang Lei is currently a Researcher with the Electric Power Research Institute of Guangdong Power Grid Co., Ltd. He received his M.S. degree in information and communication engineering from the School of Electronic Information and Communications, Huazhong University of Science and Technology in 2024. His current research interests include machine learning and applications.
Funding information
10.13039/501100001809-National Natural Science Foundation of China (Grant Number: 62575116).
Computational simulation/modeling
Not applicable
Training Robust Graph Completion Networks with Extremely Weak Supervision on Graphs with Incomplete Features and Structure
3-Aug-2026