Preview — Pro guide
You are seeing a portion of this guide. Sign in and upgrade to unlock the full article, quizzes, and interview answers.
Sections
Related Guides
Two-Stage Retrieval & Ranking: The Architecture Behind Every Large-Scale Recommender
ML System Design
Probability Calibration: When Model Scores Must Be Real Probabilities
ML System Design
Offline vs Online Evaluation: Why Metrics Disagree and What to Do About It
ML System Design
How to Design at MLSD: Blank Whiteboard to Production ML
ML System Design
MLSD Case Study: End-to-End Recommender System
ML System Design
Multi-Task Learning: Shared-Bottom, MMOE, and Negative Transfer
Every production ML ranker optimizes multiple objectives simultaneously — CTR, conversion, watch time, shares. This guide covers shared-bottom vs MMOE, negative transfer, task weighting, and the production patterns at YouTube, TikTok, Airbnb, and Meta.
Why Every Production Ranker Uses Multiple Objectives
No real product optimizes a single metric. A video recommendation ranker that optimizes click-through rate alone learns to recommend clickbait — videos with sensational thumbnails that cause users to abandon after 5 seconds. A ranker that optimizes watch time alone learns to recommend long, boring background-play content with no actual engagement.
Production ranking systems at YouTube, TikTok, Instagram, LinkedIn, and Airbnb all predict multiple outcome signals simultaneously — then combine them into a single score that reflects the full business objective. This is multi-task learning (MTL): a single model trained to predict multiple related outputs, sharing information across tasks to improve each.
The business case for MTL over separate per-task models:
- Data efficiency: related tasks share signal. A model learning CTR and conversion simultaneously benefits from conversion labels correcting CTR predictions and vice versa — sparse conversion labels are amplified by dense click labels.
- Regularization: tasks act as regularizers for each other, preventing overfit to noise in any single task's labels.
- Consistency: a single model scores items on all objectives simultaneously — no synchronization issues between N independent models serving at different latencies.
- Negative transfer awareness: multi-task architectures explicitly manage task conflict; separate models have no mechanism to reason about competing objectives.
The challenge: tasks conflict. Adding watch time as a second objective improves ranking quality for long-form content but can hurt ranking for short-form (TikTok-style) where watch-through percentage matters more than absolute duration. Task conflict causes negative transfer — adding a poorly-aligned task degrades performance on the primary task. MMOE is the architectural solution.