Enhancing missing data imputation through combined bipartite graph and complete directed graph

In this paper, we address a central challenge in tabular missing-data imputation: explicitly identifying and exploiting interdependencies among features to improve reconstruction quality. Current state-of-the-art methods mostly model similarity or interdependence between samples. However, our experiments on real-world tabular datasets show that, when samples are truly independent, building such observation-level graphs yields only marginal and dataset-specific performance gains, rather than consistent and generalizable benefits. We therefore introduce the Bipartite and Complete Directed Graph Neural Network (BCGNN). In BCGNN, observations and features are treated as two distinct node types, and each observed cell value is converted into an attributed edge connecting them. The bipartite component inductively learns node embeddings by fully leveraging the information encoded in these attributed edges, while the complete directed graph component explicitly describes and propagates intricate feature–feature dependencies. The combined graph furnishes a robust inductive framework for representation learning while explicitly parameterizing higher-order dependencies among features. Across diverse missing mechanisms, BCGNN outperforms leading imputation baselines, achieving an average 15% reduction in mean absolute error. Extensive experiments confirm that a deeper understanding of feature interdependence markedly enhances embedding quality. BCGNN also delivers superior performance on downstream label-prediction tasks with missing inputs and demonstrates robust generalization to unseen data.