pipette
ESEspañol

Learned Cross-Task Relationships in Multi-Task Models

Victor Zhang, Yiping Yuan, Florian Raudies, Bosun Adeoti, Brian Y. C. Leung, Sanjay Surendranath Girija, Naijing Zhang

Preprint with a published versionReal-world use

In the authors' words

We propose a framework that learns cross-task relationships in multi-task models by approximating the joint distribution of task labels through targeted pairwise relationships. This approach improves performance via transfer learning and enhances information extraction without the intractable complexity of modeling the full joint space. Although our framework applies to any multi-task system, we demonstrate its efficacy within YouTube's production recommendation systems. Experiments across the Notifications, Homepage, and Watch Next surfaces show improvements in both accuracy and user satisfaction metrics. Finally, we propose a workflow template to facilitate broader future implementation.

Main resultLimitation the authors admit

Appeared: Friday, September 25. arXiv. Preprint with a published version.

DOI: 10.1145/3773078.3831878

Published version: Proceedings of the 20th ACM Conference on Recommender Systems (RecSys 2026)

Authors' comment: Accepted to the 20th ACM Conference on Recommender Systems (RecSys 2026). 10 pages, 6 figures