RelPrism: A Multi-Faceted Pre-training Framework with Self-Generated Tasks for Relational Databases

Researchers introduced RelPrism, a self-supervised pre-training framework for relational databases.

RelPrism builds intrinsic, relational, and hybrid views, then clusters them at multiple granularities to create pseudo-task pools.

Models pre-trained on these tasks show improved performance on downstream classification and regression benchmarks.

The paper reports gains across 14 tasks on 5 datasets, highlighting a stronger way to learn from relational data.