Grables: Tabular Learning Beyond Independent Rows

TL;DR

This paper introduces Grables, a modular framework that enhances tabular learning by explicitly modeling inter-row dependencies through graph-based methods, improving performance on relational and temporal data.

Contribution

The paper proposes Grables, a modular interface separating graph construction from prediction, enabling better modeling of relational structures in tabular data.

Findings

Message passing captures inter-row dependencies missed by local models.

Hybrid approaches combining structure extraction with tabular learners improve accuracy.

Experiments confirm the importance of modeling relational patterns in various datasets.

Abstract

Tabular learning is still dominated by row-wise predictors that score each row independently, which fits i.i.d. benchmarks but fails on transactional, temporal, and relational tables where labels depend on other rows. We show that row-wise prediction rules out natural targets driven by global counts, overlaps, and relational patterns. To make "using structure" precise across architectures, we introduce grables: a modular interface that separates how a table is lifted to a graph (constructor) from how predictions are computed on that graph (node predictor), pinpointing where expressive power comes from. Experiments on synthetic tasks, transaction data, and a RelBench clinical-trials dataset confirm the predicted separations: message passing captures inter-row dependencies that row-local models miss, and hybrid approaches that explicitly extract inter-row structure and feed it to strong tabular learners yield consistent gains.