Self-supervised Guided Hypergraph Feature Propagation for Semi-supervised Classification with Missing Node Features

Chengxiang Lei, Sichao Fu, Yuetian Wang, Wenhao Qiu, Yachen Hu, Qinmu Peng, Xinge You

Abstract

Graph neural networks (GNNs) have demonstrated significant success in modeling networked data, yet their performance is often compromised by missing node features—a common occurrence in real-world datasets. While existing methods attempt to reconstruct these features through information propagation, they often fail to capture the complex, higher-order correlations between nodes. This paper introduces Self-supervised Guided Hypergraph Feature Propagation (SGHFP), a novel framework that leverages hypergraph structures to model multi-node relationships. By iteratively fusing a feature-based hypergraph with a self-supervised pseudo-label hypergraph, SGHFP effectively reconstructs missing data. Experimental results across multiple benchmarks demonstrate that SGHFP maintains high classification accuracy even when up to 99% of node features are missing, significantly outperforming existing state-of-the-art methods.

Introduction

Graph Neural Networks (GNNs) are the standard for analyzing data with relational structures, such as social networks, citation indices, and molecular graphs. However, a significant "Achilles' heel" for these models is the assumption of feature completeness. In practical applications, node attributes are frequently missing due to privacy concerns, sensor failures, or incomplete data collection.

Traditional approaches to this problem often rely on simple Feature Propagation (FP), which utilizes the graph's adjacency matrix to smooth features across neighboring nodes. However, these methods are limited by their pairwise nature—they only consider connections between two nodes at a time. This paper argues that such simplistic modeling fails to capture the higher-order group dynamics inherent in complex data. To address this, we propose SGHFP, which utilizes hypergraphs to better preserve data correlations during the feature reconstruction process.

Methodology

The SGHFP framework shifts the reconstruction paradigm from standard graphs to hypergraphs, where a single hyperedge can connect an arbitrary number of nodes. The process is defined by an iterative, three-stage cycle:

1. Dual-Hypergraph Construction and Fusion

SGHFP constructs two distinct hypergraphs to capture different facets of node similarity:

  • Feature Hypergraph ($G_f$): This is built using the currently available and partially reconstructed node features, grouping nodes that share similar attribute profiles.
  • Pseudo-label Hypergraph ($G_{pl}$): To incorporate semantic information, the model passes current features through a two-layer GNN to generate "pseudo-labels." Nodes predicted to belong to the same class are linked via hyperedges.

These two structures are fused into a single propagation matrix ($\Theta$), which provides a comprehensive representation of the local geometric distribution of the data.

2. Iterative Feature Propagation

Using the fused hypergraph, the model propagates known feature information to fill the gaps. This is mathematically formulated as the minimization of Dirichlet energy, ensuring that the reconstructed features vary smoothly across the hypergraph structure. This allows the model to "triangulate" missing values based on the collective information of a node's hyperedge neighbors.

3. Clamping and Refinement

To maintain fidelity to the original data, a "clamping" step is performed after each iteration. The known features are reset to their ground-truth values, preventing the model from drifting or over-smoothing the original information. This iterative refinement continues until the feature set stabilizes, at which point the reconstructed features are used for the final downstream classification task.

Experimental Evaluation

The researchers evaluated SGHFP on four benchmark datasets: Cora, CiteSeer, PubMed, and Photo. To test the model's limits, the team simulated extreme data scarcity, with missing feature rates of 50%, 90%, and 99%.

Key Results

  • Robustness to Scarcity: SGHFP demonstrated remarkable resilience. On the Cora dataset, even with 99% of features missing, the model achieved 79.41% accuracy. This is only a marginal decrease from the 81.23% accuracy achieved with 0% missing features, suggesting that the hypergraph structure successfully recovers the latent information necessary for classification.
  • Superiority over Pairwise Methods: SGHFP consistently outperformed standard Feature Propagation (FP) and other specialized methods like GCNMF and PaGNN.
  • Feature Quality: t-SNE visualizations indicated that SGHFP produces more distinct and tighter clusters of node classes compared to traditional methods, confirming that the reconstructed features are not just mathematically smooth but semantically meaningful.

Conclusion and Future Work

SGHFP provides a robust solution for GNN applications where data is incomplete. By moving beyond pairwise interactions and utilizing self-supervised pseudo-labels to guide hypergraph construction, the model effectively "fills the gaps" in high-dimensional data.

Currently, SGHFP is optimized for homogeneous graphs (where all nodes and edges are of the same type). Future research will focus on extending this hypergraph propagation framework to heterogeneous networks, which involve multiple node and edge types, potentially broadening the applicability of SGHFP to even more complex real-world scenarios.

Summary updated

Abstract

Graph neural networks (GNNs) with missing node features have recently received increasing interest. Such missing node features seriously hurt the performance of the existing GNNs. Some recent methods have been proposed to reconstruct the missing node features by the information propagation among nodes with known and unknown attributes. Although these methods have achieved superior performance, how to exactly exploit the complex data correlations among nodes to reconstruct missing node features is still a great challenge. To solve the above problem, we propose a self-supervised guided hypergraph feature propagation (SGHFP). Specifically, the feature hypergraph is first generated according to the node features with missing information. And then, the reconstructed node features produced by the previous iteration are fed to a two-layer GNNs to construct a pseudo-label hypergraph. Before each iteration, the constructed feature hypergraph and pseudo-label hypergraph are fused effectively, which can better preserve the higher-order data correlations among nodes. After then, we apply the fused hypergraph to the feature propagation for reconstructing missing features. Finally, the reconstructed node features by multi-iteration optimization are applied to the downstream semi-supervised classification task. Extensive experiments demonstrate that the proposed SGHFP outperforms the existing semi-supervised classification with missing node feature methods.
BibTeX
@inproceedings{lei2023self,
  author = {Chengxiang Lei and Sichao Fu and Yuetian Wang and Wenhao Qiu and Yachen Hu and Qinmu Peng and Xinge You},
  title = {{Self-supervised Guided Hypergraph Feature Propagation for Semi-supervised Classification with Missing Node Features}},
  booktitle = {IEEE International Conference on Acoustics, Speech and Signal Processing},
  year = {2023},
  url = {https://openreview.net/forum?id=uGS75qHQuFB},
}
Figure