Design Machine Learning Model for Facebook Groups Post Ranking
Overview
This question evaluates a candidate's ability to design an end-to-end machine learning recommender for social community group posts, probing competencies in feature engineering, label and objective selection, model family choice, offline and online evaluation, and deployment concerns such as latency, privacy, integrity, exploration, and blending with other feed content, within the Machine Learning domain for a Data Scientist role. It is commonly asked to assess trade-offs between maximizing user engagement and maintaining healthy, safe content and system constraints, and it tests both conceptual understanding of design trade-offs and practical application in evaluation, monitoring, and production deployment.
ML System Design: Ranking Facebook Groups Posts in News Feed
Scenario
You are designing a machine-learning system to recommend posts from Facebook Groups in a user's News Feed. The goal is to maximize user value and healthy engagement while respecting integrity, privacy, and latency constraints.
Assume:
- Posts can come from groups the user has joined and (optionally) public groups suggested to the user.
- The recommender will operate within a larger News Feed system and must blend well with other content types.
- Strict guardrails against spam/low-quality content and negative feedback are required.
Task
Propose a machine-learning approach covering:
- Features: What user, group, content, and context features would you use?
- Labels/Objectives: What labels and objective(s) would you optimize? How would you handle multiple signals and negative feedback?
- Model Family: What modeling architecture(s) would you choose and why?
- Offline Evaluation: Which metrics and methodologies would you use to validate offline?
- Online Evaluation: How would you run A/B tests and guardrail metrics?
- Deployment Considerations: Data/feature pipelines, latency, cold start, integrity, monitoring, exploration, and blending with the broader feed.
Hint: Consider engagement signals, embeddings, multi-task learning, and A/B validation.
Solution
Login required