Safe-Feed MMoE Recommender
Multi-task learning with Mixture-of-Experts routing
The Problem
Content recommender systems often need to optimize for multiple competing objectives simultaneously — engagement, safety, and relevance. Standard shared-bottom architectures suffer from negative transfer, where optimizing one task degrades another.
Approach
I first deployed a Shared Bottom Network to quantify the degree of negative transfer between tasks. Then I architected a Multi-gate Mixture-of-Experts (MMoE) model that uses task-specific gating networks to route inputs to different expert sub-networks, allowing each task to learn its own combination of shared representations.
Results
- Shared Bottom Network learns good features when tasks are complementary
- Multi Task Loss effectively balanced competing gradient signals