Katkat
Back

Design a newsfeed dislike model

MetaMachine Learning Engineer · Onsite
MediumMachine Learning EngineerOnsiteML System Design

Last updated

This problem tests the ability to design an end-to-end ML system for predicting user dissatisfaction, focusing on label definition, feature engineering, and the integration of negative signals into a ranking pipeline.

Problem Statement

Design a production ML system that predicts the probability of a user clicking "Dislike" or "Hide" on a social media newsfeed post. The goal is to use these predictions to down-rank content that creates a poor user experience.

Requirements

  1. Label Definition: Define what constitutes a "dislike" and how to handle implicit negative signals (e.g., hiding a post, reporting, or quickly scrolling past).
  2. Feature Engineering: Propose features that help distinguish between "content I don't like" and "content I find offensive/inappropriate."
  3. Model Architecture: Select a model type and explain how to handle the high class imbalance (dislikes are rare compared to views).
  4. System Integration: Explain how this model score should be integrated into an existing ranking system that is currently optimizing for engagement (likes, comments, time spent).
  5. Evaluation: Define offline metrics (besides AUC/LogLoss) and online metrics to monitor the success of this model.

Scale Assumptions

  • Daily Active Users (DAU): 500 million.
  • Feed request latency budget: 100ms.
  • Negative feedback events are sparse relative to total impressions.

Deliverables

  • Describe the training data pipeline (data windowing and negative sampling).
  • Explain the serving architecture: how do you score items without bloating the latency of the feed generation?
  • How do you address the 'cold start' problem for new posts?

Hints

Hint 1

Consider the difference between a direct user action (like/dislike) and an inferred signal (scrolling speed), and how you might weight these labels differently.

Hint 2

Think about the ranking objective: should you subtract the dislike score from the engagement score, or treat it as a separate constraint/threshold?

Hint 3

Address the class imbalance by discussing techniques like oversampling the minority class or using cost-sensitive learning.

Follow-up questions

  • How would you measure the trade-off between increasing user satisfaction (long-term retention) and potentially decreasing short-term engagement (e.g., click-through rates)?
  • If a specific creator receives a high volume of dislikes, how should the system differentiate between 'bad content' and 'brigading' or malicious reporting?
  • How would you handle model drift if users' content preferences change significantly during a holiday or major global event?