Katkat
Back

Design a Personalized Feed and Compare Ranking Objectives

LinkedInMachine Learning Engineer · Technical Screen
MediumMachine Learning EngineerTechnical ScreenML System Design

Last updated

Design an end-to-end personalized professional feed system and compare pointwise, pairwise, and listwise learning-to-rank formulations across offline training, online serving, and business metrics.

Scenario

You are tasked with designing the machine learning system powering the home feed for a professional social network. The feed delivers a personalized stream of content, including posts from direct professional connections, followed companies, industry influencers, recommended articles, and sponsored updates.

Your design must cover the end-to-end workflow from candidate retrieval to final presentation, with a deep dive into learning-to-rank (LTR) objective formulation.

Requirements & Scale

  • User Base: 200 million daily active users (DAU).
  • Throughput: Average 30,000 requests per second (QPS), peaking at 70,000 QPS.
  • Latency Budget: Total end-to-end response time must be under 150 ms (p99).
  • Engagement Signals: Implicit feedback (dwell time, clicks, impressions) and explicit feedback (reactions, comments, shares, saves).

Design Expectations

  1. End-to-End Architecture

    • Detail the stages from candidate generation (retrieval) to coarse ranking, fine ranking, and re-ranking (diversity, freshness, business logic).
    • Describe the data pipelines required for real-time feature extraction vs. batch offline feature computation.

  2. Feature Engineering

    • Outline key feature groups: user-level features (e.g., industry, seniority), author/content features (e.g., text embeddings, topic tags), graph-based features (e.g., degrees of connection, interaction history), and real-time contextual features.

  3. Ranking Objectives: Pointwise vs. Pairwise vs. Listwise

    • Pointwise: Formulate binary classification or regression targets (e.g., predicting P(click), P(comment), expected dwell time). Discuss calibration and multi-task learning.
    • Pairwise: Explain formulations like RankNet or LambdaMART using item pairs (post_i > post_j) for a given user session.
    • Listwise: Formulate loss over the full list (e.g., ListNet, LambdaLoss) directly optimizing metrics like NDCG.
    • Comparison: Systematically compare the three approaches across training complexity, inference efficiency, ability to handle position bias, and alignment with multi-objective business goals.

  4. Online Serving & Latency Mitigation

    • Explain how model inference fits within the 150 ms budget (e.g., two-pass scoring, vector search ANN, feature caching, model distillation).

Hints

Hint 1

Structure the candidate pipeline into at least three distinct stages (retrieval, scoring, re-ranking) to prune millions of posts down to the top 20 within the latency budget.

Hint 2

When comparing LTR paradigms, consider both mathematical objective alignment with metrics like NDCG and practical constraints such as training data generation and serving cost.

Hint 3

A production feed typically optimizes for multiple actions simultaneously; consider how a multi-task pointwise model compares in operational simplicity to pairwise or listwise alternatives.

Follow-up questions

  • How do you identify and mitigate position bias and presentation bias in your training logs?
  • How would you integrate sponsored content into the organic feed while preserving user experience and maximizing revenue?
  • If your objective shifts from maximizing short-term engagement to fostering long-term content creator retention, what architectural and objective changes would you make?