From Rows to Relations: Graph Neural Networks for Tabular Data Learning

Abstract

Tabular data remain the backbone of scientific discovery, industrial decision-making, and financial modeling, yet the methods that dominate—most notably gradient-boosted trees—still treat tables as isolated rows of features. This perspective overlooks a crucial truth: even in "flat" data, hidden relational structures exist. Instances may cluster, features may correlate, and missing values may reveal patterns of dependence. Such latent structure, if uncovered, can dramatically improve prediction, generalization, and interpretability. Graph machine learning offers exactly this lens: it transforms a table into a network of instances and features, enabling the flow of information across related entities and embedding inductive biases that trees and transformers alone cannot capture. In this talk, I will survey four converging directions that exemplify this paradigm. We begin with graph structure learning for tables (TabGSL), dynamically inducing edges to uncover implicit correlations. We then explore bipartite graph neural networks (GTab), which enable holistic learning, zero-shot transfer across divergent schemas, and feature-incremental inference. Next, we examine missingness-aware GNNs (MissGNN), which unify imputation and prediction in a single relational framework. Finally, we highlight graph-regularized boosting (gbtGNN), where trees and GNNs co-train in a synergistic loop, setting new benchmarks across both full-data and few-shot regimes. Together these efforts chart a new research frontier: reimagining tabular learning not as isolated row-wise classification, but as relational reasoning over learned graphs. By embracing this view, we unlock models that are more adaptable, robust, and generalizable, pointing toward the next generation of tabular AI.