New Two-Stage Framework Boosts Graph Feature Imputation Under Extreme Sparsity

Global: New Two-Stage Framework Boosts Graph Feature Imputation Under Extreme Sparsity

Researchers have introduced a novel approach to tackle the persistent problem of missing node attributes in graph‑structured data. The method, named FSD-CAP, was detailed in a recent arXiv preprint and aims to maintain predictive performance even when 99.5% of features are absent. By targeting both structural and uniform missing patterns, the framework seeks to reduce error propagation that commonly afflicts existing diffusion‑based techniques.

Method Overview

FSD-CAP operates in two distinct stages. The initial stage confines diffusion to locally relevant subgraphs, leveraging graph‑distance metrics to expand neighborhoods selectively. A fractional diffusion operator then modulates the sharpness of information spread, adapting to the density of connections in each region of the graph.

Stage One: Localized Diffusion

During subgraph expansion, nodes are grouped based on proximity, ensuring that only structurally pertinent vertices participate in the diffusion process. The fractional diffusion component dynamically adjusts the propagation coefficient, which helps prevent the amplification of noise that can arise when global diffusion is applied to highly sparse inputs.

Stage Two: Class‑Aware Refinement

The second stage refines the imputed features through class‑aware propagation. Pseudo‑labels generated from an interim classifier guide the diffusion, while neighborhood entropy metrics encourage consistency among adjacent nodes. This combination aims to align the reconstructed features with the underlying class distribution.

Experimental Evaluation

The authors tested FSD-CAP on five benchmark datasets, each subjected to a 99.5% feature‑missing scenario. Both structural and uniform missing regimes were examined, providing a comprehensive view of the framework’s robustness across diverse graph topologies.

Node Classification Performance

Under the extreme sparsity condition, FSD-CAP achieved average node‑classification accuracies of 80.06% for structurally missing features and 81.01% for uniformly missing features. These results approach the 81.31% accuracy recorded by a standard Graph Convolutional Network (GCN) when full feature information is available.

Link Prediction Results

For link‑prediction tasks, the framework recorded Area Under the Curve (AUC) scores of 91.65% (structural) and 92.41% (uniform). By comparison, the fully observed baseline reached 95.06% AUC, indicating that FSD‑CAP narrows the performance gap considerably despite the severe data loss.

Comparative Analysis and Implications

Across all experiments, FSD‑CAP outperformed competing imputation models, particularly on large‑scale graphs and datasets characterized by heterophily. The ability to retain high classification and link‑prediction performance suggests potential applicability in domains such as social network analysis, recommender systems, and biological network modeling, where missing attribute information is common.

Future Directions

The authors note that extending the framework to dynamic graphs and exploring alternative pseudo‑labeling strategies could further enhance resilience to missing data. Ongoing work aims to integrate the approach with downstream tasks beyond classification and link prediction.

This report is based on information from arXiv, licensed under Academic Preprint / Open Access. Based on the abstract of the research paper. Full text available via ArXiv.