Graph Attribute Imputation via Generalized Wasserstein Balancing

ICLR 2026 Conference Submission17794 Authors

Keywords: Wasserstein Balancing, Graph Attribute Imputation

Abstract: Distribution alignment methods effectively impute missing values in tabular datasets by assuming consistent distributions across batches and minimizing discrepancies among them. However, directly applying these methods to graph data is challenging: (1) standard discrepancy measures neglect structural information, and (2) noise in graphs tends to be propagated and amplified through structural dependencies, ultimately degrading imputation performance. To address these challenges, we propose the Relaxed Graph Spectral Discrepancy (RGSD), a discrepancy designed to compare sets of graphs by capturing both structural patterns and inter-node correlations through spectral decomposition, along with a selective matching regularization to mitigate the impact of noise. Building on RGSD, we introduce the RGSD for Imputation (RGSImp) framework, which iteratively refines graph imputation results by minimizing the RGSD between observed and imputed data. Experiments on multiple benchmarks demonstrate that RGSImp effectively incorporates graph structure and node correlations, achieving superior performance over state-of-the-art graph imputation methods in both imputation accuracy and downstream tasks.

Primary Area: other topics in machine learning (i.e., none of the above)

Submission Number: 17794