The Missing Structure: When Graph Representations Outperform Tabular Models

Tamara Cucumides, Floris Geerts

Submission Type: Short paper (4 pages)

Keywords: graph learning, tabular data, message passing

TL;DR: When labels depend on other rows, traditional tabular models fail and message passing on graphs induced from the table captures the necessary structure.

Abstract: Row-local tabular models (gradient-boosted trees and MLPs) excel when labels depend only on per-row attributes. Yet many real-life labels depend on \emph{other rows} (shared values, references, group effects). We ask when explicit cross-row structure becomes \emph{necessary}. Starting from a single table, we construct controlled row-level tasks that require existence or counting, and compare (i) strong row-local learners, (ii) the same learners with one-hop neighbor feature aggregation (NFA), and (iii) message passing on graphs induced directly from the table. In this controlled setting, NFA yields only small and inconsistent gains over row-local baselines, suggesting that static neighbor summaries alone are insufficient to recover relational dependencies. Message passing reliably captures the required cross-row logic. These findings reveal a structural difference between tabular and graph learning and suggest that dynamic propagation, rather than static aggregation, is key when targets depend on other rows.

Submission Number: 31