Design a Video Recommender with Two-Tower Retrieval and MMoE Ranking
Company: Snapchat
Role: Machine Learning Engineer
Category: Machine Learning
Difficulty: medium
Interview Round: Onsite
Design a video recommendation system, including log-to-feature and label flow, two-tower retrieval updates, serving, and offline evaluation. Explain how a multi-gate mixture-of-experts model could support several ranking targets and what its losses would be.
### Constraints and Clarifications
No specific labels, update cadence, traffic target, or utility function is provided. Give one complete illustrative baseline with explicit choices, then explain how measurements would change it. Do not claim the chosen settings describe an employer's deployed system.
### What a Strong Answer Covers
Timestamped examples, label maturity, candidate retrieval, embedding/model version consistency, task-specific gates over shared experts, concrete losses and serving score, and evaluation aligned with those targets.
### Follow-up Questions
- How do you keep candidate embeddings compatible with an updated query tower?
- What happens when the tasks compete or one label is much sparser than another?
Overview: Build a concrete video recommendation baseline with two-tower retrieval and MMoE ranking. Define watch and like labels, losses, a serving utility score, compatible embedding updates, and aligned evaluation.
Read the full Snapchat Machine Learning Engineer interview experience this question came from