Context

Meta products such as Facebook Feed, Instagram Feed, and Reels rely on recommendation systems to select and order content from a very large candidate set in real time. Interviewers ask this to assess whether you can translate a broad product problem into concrete algorithmic stages.

Core Question

Walk through how a recommendation system for Facebook Feed might work from input signals to final ranked results. In your answer, address:

1. How candidate content is generated efficiently from a massive graph of users, creators, and posts.
2. How candidates are scored and ranked using features such as affinity, recency, and predicted engagement.
3. How the system handles practical issues like cold start, diversity, feedback loops, and latency constraints.

Scope Guidance

You do not need to derive a specific ML model, but you should explain the algorithmic pipeline clearly: retrieval, filtering, ranking, and post-processing. A strong answer connects data structures and algorithms—such as graph traversal, heaps for top-k selection, and sorting/reranking—to product goals and system constraints.

Candidate Retrieval

The system cannot score every possible post, so it first retrieves a manageable candidate set. This often uses graph-based expansion from friends, followed accounts, similar users, or similar content, plus lightweight heuristics to keep latency low.

candidates = set(friends_posts + followed_creator_posts + similar_user_posts)

Feature-Based Scoring

Each candidate is represented by features such as user-author affinity, content freshness, predicted click probability, watch time, or likelihood of meaningful interaction. A ranking model combines these signals into a score used to order items.

score = 0.5 * affinity + 0.3 * engagement_prob + 0.2 * freshness

Top-K Selection

After scoring, the system usually needs only the best few items for the first screen. A heap is a common way to keep the top-k candidates efficiently without fully sorting a very large list.

import heapq
best = heapq.nlargest(k, scored_items, key=lambda x: x[1])

Reranking and Constraints

Pure score ordering is rarely enough because it can over-concentrate on one creator, topic, or content type. A reranking stage applies business and quality constraints such as diversity, integrity filters, and creator caps.

if creator_count[item.creator] < limit and item.safe:
    result.append(item)

Feedback and Adaptation

Recommendations improve by learning from impressions, clicks, likes, comments, hides, and dwell time. The system updates user representations and model features over time, balancing recent behavior with longer-term preferences.

user_profile[topic] += alpha * feedback_signal