Task Scarcity and Label Leakage in Relational Transfer Learning

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Azevedo, Francisco Galuppo, Loures, Clarissa Lima, Correa, Denis Oliveira
Format: Preprint
Published: 2026
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866912992579289088
author Azevedo, Francisco Galuppo
Loures, Clarissa Lima
Correa, Denis Oliveira
author_facet Azevedo, Francisco Galuppo
Loures, Clarissa Lima
Correa, Denis Oliveira
contents Training relational foundation models requires learning representations that transfer across tasks, yet available supervision is typically limited to a small number of prediction targets per database. This task scarcity causes learned representations to encode task-specific shortcuts that degrade transfer even within the same schema, a problem we call label leakage. We study this using K-Space, a modular architecture combining frozen pretrained tabular encoders with a lightweight message-passing core. To suppress leakage, we introduce a gradient projection method that removes label-predictive directions from representation updates. On RelBench, this improves within-dataset transfer by +0.145 AUROC on average, often recovering near single-task performance. Our results suggest that limited task diversity, not just limited data, constrains relational foundation models.
format Preprint
id arxiv_https___arxiv_org_abs_2603_29914
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle Task Scarcity and Label Leakage in Relational Transfer Learning
Azevedo, Francisco Galuppo
Loures, Clarissa Lima
Correa, Denis Oliveira
Machine Learning
Training relational foundation models requires learning representations that transfer across tasks, yet available supervision is typically limited to a small number of prediction targets per database. This task scarcity causes learned representations to encode task-specific shortcuts that degrade transfer even within the same schema, a problem we call label leakage. We study this using K-Space, a modular architecture combining frozen pretrained tabular encoders with a lightweight message-passing core. To suppress leakage, we introduce a gradient projection method that removes label-predictive directions from representation updates. On RelBench, this improves within-dataset transfer by +0.145 AUROC on average, often recovering near single-task performance. Our results suggest that limited task diversity, not just limited data, constrains relational foundation models.
title Task Scarcity and Label Leakage in Relational Transfer Learning
topic Machine Learning
url https://arxiv.org/abs/2603.29914