Meta·Data Scientist·Technical Phone Screen·Senior

SeniorPrefer not to say

Jun 2026Remote

Summary

Meta DS interview with a meaty system design question about building a hashtag recommendation system. One question but it had like four parts baked in, so it felt more like a 30-minute conversation than a single prompt.

Questions Asked (1)

Design a hashtag recommendation system for a social media platform. What signals and features would you collect? How would you handle cold-start users with little or no data? How would you combine those features into a scoring function, and how would you learn or set the weights?

System DesignProduct Analytics & MetricsTechnical Trade-offs

Author's notes

The cold-start part is where I stumbled.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by clarifying the product context and success metrics, then outline a multi-signal feature set (content, user, context, and social graph). Address cold-start with content-based and heuristic fallbacks, and describe a hybrid scoring function with learned weights via online learning and offline evaluation.

Pro tip: Emphasize that the system should optimize for long-term user engagement and diversity, not just immediate click-through rate, and discuss how you'd measure and mitigate feedback loops and popularity bias.

1. Clarify Requirements and Metrics

Ask about the platform's scale, user base, and business goals. Define success metrics such as CTR, hashtag adoption rate, and session depth, and consider trade-offs between relevance and diversity.

2. Identify Signals and Features

List features from four categories: content (post text, image, hashtag embeddings), user (history, demographics, interests), context (time, location, device), and social (trending, friend usage). Explain how each signal could improve recommendations.

3. Handle Cold-Start Users

For new users, use content-based filtering (e.g., recommend hashtags similar to post content), demographic-based popularity, and contextual signals. For new hashtags, use content similarity and early engagement signals to bootstrap.

4. Design Scoring Function and Learning

Combine features into a weighted score (e.g., linear model or gradient-boosted trees). Learn weights via online learning (e.g., multi-armed bandits) and offline training on logged data, with exploration to avoid feedback loops.

5. Evaluate and Iterate

Use offline metrics (precision@k, NDCG) and online A/B tests to measure engagement and diversity. Monitor for bias and adjust the model periodically.

Key Points to Mention

Use a hybrid recommendation approach combining collaborative filtering and content-based methods.
Incorporate real-time trending hashtags and social proof (e.g., friends' usage) as signals.
For cold-start, leverage post content (text, image) and user context (location, time) to make initial recommendations.
Model the problem as a ranking task and use learning-to-rank algorithms.
Address the exploration-exploitation trade-off with multi-armed bandits or Thompson sampling.
Evaluate beyond accuracy: consider diversity, novelty, and long-term user satisfaction.

AI-generated suggestions, not part of the candidate's original notes. May be inaccurate — verify before relying on them.