Framework Learns Cross-Task Relationships for Multi-Task Models

Summary

The paper proposes a framework for learning cross-task relationships in multi-task models by approximating the joint distribution of task labels through targeted pairwise relationships. The approach is intended to enable transfer learning and better information extraction without the intractable complexity of modeling the full joint space. The authors evaluate it in YouTube’s production recommendation systems across the Notifications, Homepage, and Watch Next surfaces. They report improvements in both accuracy and user-satisfaction metrics. The paper also presents a workflow template for implementing the approach in other multi-task systems. It has been accepted to the 20th ACM Conference on Recommender Systems (RecSys 2026).