← Airbnb Interview Insights

Airbnb·Machine Learning Engineer·Onsite - System Design / Architecture·Senior

SeniorPrefer not to say
Jun 2026

Summary

ML system design interview at Airbnb for an MLE role, basically one long deep-dive into building a search ranking system from scratch. Pretty thorough coverage of the whole stack, from retrieval pipelines to fairness concerns to A/B testing.

Questions Asked (9)

Q1

How would you frame the search ranking problem for Airbnb? Should you use pointwise, pairwise, or listwise learning-to-rank, and what label would you use (booking, click, or something else)?

System DesignTechnical Trade-offsProduct Analytics & Metrics
Author's notes

This is where I spent probably too long.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by clarifying the business objective and the two-sided marketplace dynamics (guest bookings vs. host exposure). Then compare pointwise, pairwise, and listwise LTR in terms of data requirements, computational cost, and alignment with Airbnb's ranking metrics. Finally, recommend a label that balances business value and data availability, such as booking with click as an auxiliary signal.

Pro tip: Emphasize that Airbnb's ranking is a multi-objective problem: you need to optimize for bookings while ensuring fairness and diversity across hosts. Mention that using booking as the primary label with click as a secondary signal can mitigate position bias and sparsity.

1. Clarify business goals and constraints

Identify key objectives: maximize bookings, revenue, guest satisfaction, and host fairness. Consider constraints like latency, scalability, and cold-start.

2. Compare LTR approaches

Discuss pointwise (independent prediction), pairwise (relative order), and listwise (optimize entire list) in terms of data, complexity, and suitability for Airbnb's search.

3. Choose a label

Evaluate booking, click, and other signals (e.g., dwell time, favorites) as labels. Consider using booking as primary and click as auxiliary to handle sparsity and position bias.

4. Address practical challenges

Mention position bias, feedback loops, and evaluation metrics (NDCG, MRR). Suggest debiasing techniques like inverse propensity scoring.

5. Recommend a solution

Propose a hybrid approach: use listwise LTR with booking as the main label, augmented with click data, and incorporate business rules for diversity and fairness.

Key Points to Mention

  • Pointwise, pairwise, and listwise LTR trade-offs: pointwise is simple but ignores ranking; pairwise models relative order; listwise optimizes the entire list but is data-hungry.
  • Booking as the ultimate business metric, but it's sparse and delayed; click is abundant but noisy and biased.
  • Position bias in click data and how to mitigate it (e.g., inverse propensity scoring, click models).
  • Two-sided marketplace: need to balance guest relevance with host exposure and fairness.
  • Evaluation metrics: NDCG, MRR, and online A/B testing with business KPIs.
  • Cold-start and exploration: new listings need exposure, so consider exploration strategies.

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

Q2

Walk me through the data sources you'd use to train the ranking model. What features come from listings, hosts, users, and the query context itself?

System DesignData Modeling
Author's notes

Felt okay here.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by framing the ranking problem as a two-sided marketplace where you need to model both guest preferences and host quality. Then systematically walk through each data source (listings, hosts, users, query context) and map them to feature groups, explaining how they interact in the ranking model. Conclude by discussing how you'd handle data quality, leakage, and real-time serving constraints.

Pro tip: Emphasize that query context features (e.g., search dates, party size, location) are often the strongest signals because they capture immediate intent, and mention that you'd validate feature importance with offline metrics like NDCG before deploying.

1. Clarify the ranking objective and data sources

State that the goal is to rank listings by likelihood of booking or guest satisfaction, and list the primary data sources: listings, hosts, users, and query context. Mention that you'd also incorporate historical interaction data (clicks, bookings) as labels.

2. Enumerate listing features

Describe features derived from listings: property type, room type, amenities, price, location (lat/long, neighborhood), photos, description text (via NLP), and availability calendar. Note that these are static or slowly changing.

3. Enumerate host features

Cover host-level attributes: host tenure, response rate, acceptance rate, superhost status, number of listings, and historical ratings. Explain that these capture reliability and trust.

4. Enumerate user features

Discuss user-specific features: past bookings, search history, price sensitivity, preferred amenities, and demographic data (if available). Mention that these personalize the ranking.

5. Enumerate query context features

Detail query context: search dates, number of guests, destination, filters applied, device type, and time of day. Highlight that these are crucial for capturing real-time intent and should be computed at serving time.

Key Points to Mention

  • Use of embeddings for high-cardinality categorical features (e.g., listing IDs, user IDs) to handle sparsity.
  • Temporal features: seasonality, day of week, lead time, and recency of user activity.
  • Interaction features: user-listing affinity (e.g., past clicks/bookings), price difference from user's average, and location match.
  • Data leakage prevention: ensure features are computed only from data available at prediction time (e.g., no future bookings).
  • Real-time vs. batch features: distinguish between precomputed (e.g., host response rate) and on-the-fly (e.g., query context) features.
  • Evaluation metrics: offline (NDCG, MAP) and online (CTR, booking rate) to validate feature usefulness.

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

Q3

Describe the two-stage retrieval and ranking pipeline you'd build. How does the candidate retrieval stage differ from the final ranker?

System DesignTechnical Trade-offs
Author's notes

Pretty comfortable with this one.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by framing the two-stage architecture as a latency-accuracy trade-off, then contrast the retrieval stage (optimized for recall and speed) with the ranking stage (optimized for precision and personalization). Use Airbnb-specific examples like home recommendations to ground your answer in the company's context.

Pro tip: Emphasize that the retrieval stage is a coarse filter where you can afford simpler models (e.g., two-tower embeddings) and approximate nearest neighbor search, while the ranker can be a heavy model with rich features—this shows you understand production constraints.

1. Define the two-stage pipeline

Explain that the retrieval stage quickly narrows down millions of items to a few hundred candidates, and the ranking stage then precisely orders those candidates for the user.

2. Describe candidate retrieval

Detail how retrieval uses lightweight models (e.g., two-tower neural networks, matrix factorization) and approximate nearest neighbor search (e.g., FAISS, ScaNN) to optimize for recall and low latency.

3. Describe the final ranker

Explain that the ranker uses a more complex model (e.g., gradient boosted trees, deep neural networks) with rich features (user, item, context, cross features) to optimize for precision and business metrics.

4. Contrast the stages

Highlight differences in objectives (recall vs. precision), model complexity, feature richness, latency constraints, and training data (implicit vs. explicit feedback).

5. Discuss trade-offs and Airbnb context

Mention how you'd balance latency and accuracy, handle cold start, and incorporate Airbnb-specific signals like host quality, price, and location.

Key Points to Mention

  • Recall vs. precision trade-off between stages
  • Use of approximate nearest neighbor search (e.g., FAISS) in retrieval
  • Two-tower model architecture for retrieval
  • Feature engineering and cross features in ranking
  • Latency constraints and model complexity
  • Airbnb-specific signals: host quality, price, location, user preferences

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

Q4

What features would you engineer for the ranker? Think about listing quality, price competitiveness, host responsiveness, and personalization.

System DesignData Modeling
Author's notes

I liked this part.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by framing the ranking problem: the goal is to rank listings to maximize booking likelihood and guest satisfaction. Then, for each of the four areas (listing quality, price competitiveness, host responsiveness, personalization), propose specific features, explaining how they are computed and why they matter. Finally, discuss how to combine them into a model, mentioning feature interactions and validation.

Pro tip: Emphasize that features should be designed with business metrics in mind, and that you would validate their impact through online experiments (A/B tests) rather than relying solely on offline metrics.

1. Clarify the ranking objective

Define what the ranker is optimizing for (e.g., booking conversion, guest satisfaction) and how features support that objective.

2. Brainstorm features per category

For each of the four areas, list concrete features: listing quality (ratings, photos, amenities), price competitiveness (relative price, value score), host responsiveness (response rate, time), personalization (user history, search context).

3. Explain feature engineering details

Describe how to compute each feature, including data sources, transformations (e.g., normalization, embeddings), and handling of missing values.

4. Discuss model integration and interactions

Explain how features are combined (e.g., in a gradient boosted tree or neural network), and highlight important interactions (e.g., price sensitivity varies by user).

5. Outline validation and iteration

Mention offline evaluation (e.g., NDCG, AUC) and online A/B testing to measure feature impact, and how to iterate based on results.

Key Points to Mention

  • Listing quality: use review scores, number of reviews, photo quality, amenity completeness, and freshness of listing.
  • Price competitiveness: compute relative price compared to similar listings in the same area and dates, and consider value-for-money scores.
  • Host responsiveness: include response rate, average response time, acceptance rate, and cancellation rate.
  • Personalization: incorporate user's past bookings, search history, click behavior, and contextual signals (e.g., device, trip purpose).
  • Feature interactions: e.g., price sensitivity interacts with user income or past booking price range; responsiveness matters more for last-minute bookings.
  • Validation: use offline metrics like NDCG and online A/B tests to measure impact on bookings and guest satisfaction.

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

Q5

What model architecture would you choose for the ranker, and why? Compare gradient boosted trees versus a neural ranker.

Technical Trade-offsSystem Design
Author's notes

GBDT for interpretability and solid baseline, DNN if you want to incorporate embeddings and handle sparse features better.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by clarifying the ranking context (e.g., search, recommendations) and the key constraints (latency, data volume, feature complexity). Then compare GBTs and neural rankers across those dimensions, and recommend a hybrid or staged approach that balances performance and practicality.

Pro tip: Emphasize that the best model depends on the specific ranking stage and business metrics; mention that Airbnb often uses a two-stage system with a lightweight GBT for initial ranking and a neural model for final re-ranking, which shows you understand real-world production trade-offs.

1. Clarify the ranking problem

Ask about the specific use case (e.g., search ranking, recommendation), data scale, latency requirements, and available features. This ensures your answer is tailored to the context.

2. Compare GBTs and neural rankers

Discuss strengths and weaknesses: GBTs excel with tabular data, are interpretable, and train fast; neural rankers handle complex feature interactions, embeddings, and large-scale data but require more tuning and infrastructure.

3. Evaluate trade-offs for the given context

Map the model characteristics to the constraints from step 1. For example, if latency is critical and features are mostly tabular, GBTs may be preferable; if rich user/item embeddings and sequential behavior are key, neural rankers win.

4. Propose a recommendation

State your choice clearly, justifying it with the trade-offs. Consider suggesting a hybrid approach (e.g., GBT for initial ranking, neural for re-ranking) if it fits the scenario.

5. Discuss evaluation and iteration

Mention how you would evaluate the model (offline metrics like NDCG, online A/B tests) and iterate, showing a production mindset.

Key Points to Mention

  • GBDTs: strong for tabular data, feature importance, fast training, but limited in handling high-cardinality categorical features and complex interactions without manual feature engineering.
  • Neural rankers: can learn embeddings, handle sparse features, and model sequential/user-item interactions, but require large data, careful regularization, and more compute for training and serving.
  • Latency and infrastructure: GBDTs often have lower inference latency and simpler deployment; neural models may need GPUs and more complex serving stacks.
  • Hybrid approaches: use GBDT for candidate generation or initial ranking, then neural for final ranking to balance quality and efficiency.
  • Evaluation metrics: NDCG, MAP, recall@k for offline; online metrics like CTR, conversion, and revenue per search.
  • Airbnb context: mention their use of neural networks for embeddings (e.g., listing embeddings) and GBDTs for ranking in production systems.

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

Q6

What offline and online metrics would you use to evaluate the ranking system?

Product Analytics & MetricsA/B Testing & Experimentation
Author's notes

NDCG and AUC offline, booking rate and search-to-booking conversion online.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by clarifying the ranking system's objective (e.g., relevance, booking conversion) and then structure your answer around offline metrics (e.g., NDCG, recall) and online metrics (e.g., CTR, conversion rate). Emphasize the importance of aligning offline metrics with online business outcomes and using A/B testing to validate improvements.

Pro tip: Highlight the trade-off between offline and online metrics: offline metrics are fast and cheap but may not capture user behavior, while online metrics are the ultimate test but require careful experiment design. Mention guardrail metrics to ensure long-term health.

1. Clarify the ranking objective

Ask clarifying questions to understand what the ranking system is optimizing for (e.g., guest booking likelihood, host quality, long-term satisfaction). This ensures your metrics align with business goals.

2. Define offline metrics

List offline evaluation metrics such as NDCG, MAP, MRR, precision@k, recall@k, and AUC. Explain how they measure ranking quality using historical or labeled data.

3. Define online metrics

Describe online metrics like click-through rate (CTR), conversion rate, booking rate, revenue per user, and engagement metrics. These measure real user interactions in live experiments.

4. Connect offline to online

Explain how offline metrics can predict online performance, but note that they are not perfect. Mention techniques like counterfactual evaluation or offline-online correlation analysis.

5. Include guardrail and long-term metrics

Discuss guardrail metrics (e.g., latency, diversity, fairness) and long-term metrics (e.g., repeat bookings, host retention) to ensure the ranking system doesn't harm user experience or business health.

Key Points to Mention

  • Offline metrics: NDCG, MAP, MRR, precision@k, recall@k, AUC
  • Online metrics: CTR, conversion rate, booking rate, revenue per user, engagement
  • A/B testing and experiment design (e.g., randomization, sample size, statistical significance)
  • Guardrail metrics: latency, diversity, fairness, and system health
  • Long-term metrics: repeat bookings, host retention, user satisfaction
  • Trade-offs between offline and online metrics, and the importance of aligning with business objectives

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

Q7

How would you handle cold-start for new listings and new users who have no historical data?

System DesignTechnical Trade-offsAdaptability & Ambiguity
Author's notes

Blanked for a second on new listings specifically.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by clarifying the specific cold-start scenario (new listing vs. new user) and the business goals (e.g., ranking, recommendations). Then outline a multi-pronged strategy that leverages content-based features, transfer learning, and exploration techniques, while addressing evaluation challenges.

Pro tip: Emphasize the importance of defining a clear fallback strategy and a smooth transition from cold-start to warm-start as data accumulates. Also, discuss how you would measure success and iterate quickly.

1. Clarify the problem and constraints

Ask clarifying questions to understand the specific cold-start scenario, available data, and business objectives. Identify whether the focus is on new listings, new users, or both, and what metrics matter.

2. Leverage content and metadata

Use available attributes (e.g., listing descriptions, photos, user demographics) to build content-based models that can make initial predictions without interaction data.

3. Apply transfer learning and meta-learning

Utilize models pre-trained on similar tasks or domains to initialize predictions for new entities. Consider meta-learning approaches that learn to adapt quickly from few examples.

4. Incorporate exploration and active learning

Design bandit-based or active learning strategies to gather feedback efficiently, balancing exploitation of known preferences with exploration of new items/users.

5. Plan for evaluation and iteration

Define offline and online evaluation metrics, set up A/B tests, and create a feedback loop to transition from cold-start to warm-start as data accumulates.

Key Points to Mention

  • Content-based filtering using item/user features
  • Transfer learning from related domains or tasks
  • Meta-learning for few-shot adaptation
  • Exploration-exploitation trade-off (e.g., multi-armed bandits)
  • Hybrid models combining content and collaborative signals
  • Evaluation challenges and metrics for cold-start scenarios

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

Q8

How do you ensure fairness and result dispersion in the ranked list? What happens if the model always surfaces the same top listings?

Product StrategyTechnical Trade-offsProduct Analytics & Metrics
Author's notes

This was the most interesting part of the conversation to me.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by defining what fairness and result dispersion mean in the context of Airbnb's search ranking, then explain how you would measure and monitor them. Discuss the trade-offs between fairness, relevance, and business goals, and propose concrete techniques to ensure diversity and avoid over-concentration. Finally, address the scenario where the model always surfaces the same top listings, outlining mitigation strategies and their implications.

Pro tip: Emphasize that fairness is not just a technical metric but also a product and business decision—show that you understand the need to balance guest satisfaction, host fairness, and long-term marketplace health. Mention that you would run A/B tests to measure the impact of diversity interventions on both short-term and long-term metrics.

1. Define fairness and dispersion metrics

Clarify what fairness means for Airbnb (e.g., equitable exposure for hosts, avoiding bias against certain demographics) and how to measure result dispersion (e.g., entropy, Gini coefficient, coverage of listings).

2. Diagnose over-concentration

Analyze the ranked list to identify if the same top listings are repeatedly surfaced, using metrics like top-K frequency, unique listing count, and exposure distribution across hosts.

3. Apply mitigation techniques

Implement techniques such as diversity constraints, re-ranking, exploration-exploitation (e.g., epsilon-greedy), or fairness-aware learning to ensure a more balanced exposure while maintaining relevance.

4. Evaluate trade-offs and iterate

Assess the impact of these techniques on key metrics (e.g., booking conversion, guest satisfaction, host retention) through A/B tests, and iterate to find the optimal balance.

5. Monitor and govern

Set up ongoing monitoring and governance to detect fairness issues and result concentration, and establish a cross-functional process to address them proactively.

Key Points to Mention

  • Fairness definitions: individual vs. group fairness, and how they apply to two-sided marketplaces like Airbnb.
  • Metrics for dispersion: entropy, Gini coefficient, coverage, and top-K diversity.
  • Techniques: diversity constraints, re-ranking, exploration-exploitation, fairness-aware learning to rank.
  • Trade-offs: relevance vs. fairness, short-term vs. long-term business impact.
  • A/B testing and causal inference to measure the effect of fairness interventions.
  • Business implications: host satisfaction, regulatory compliance, and brand trust.

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

Q9

How would you handle online serving given latency constraints, and how would you design an A/B test to validate the new ranker?

A/B Testing & ExperimentationSystem Design
Author's notes

Standard stuff but easy to mess up the details.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by outlining the latency constraints for online serving (e.g., p99 < 100ms) and describe a two-stage architecture with candidate generation and ranking, using techniques like model quantization and caching. Then, for A/B testing, define clear metrics (e.g., booking rate, revenue), design a randomized controlled experiment with sufficient power, and discuss guardrail metrics and long-term effects.

Pro tip: Emphasize the trade-off between model complexity and latency, and propose a shadow deployment or interleaving test before full A/B test to catch issues early. Also, mention the importance of analyzing heterogeneous treatment effects to understand impact across user segments.

1. Clarify Requirements and Constraints

Ask about latency SLAs (e.g., p99 < 100ms), throughput, and business goals. Confirm the current serving architecture and pain points.

2. Design Low-Latency Serving Architecture

Propose a two-stage system: candidate generation (e.g., ANN) and ranking (e.g., GBDT/NN). Discuss optimizations like model quantization, pruning, caching, and precomputation.

3. Define A/B Test Objectives and Metrics

Identify primary metric (e.g., bookings per user) and guardrail metrics (e.g., latency, cancellation rate). Ensure metric sensitivity and alignment with business goals.

4. Design Experiment Setup

Randomize at user level, determine sample size and duration via power analysis, and consider stratification. Plan for novelty effects and long-term holdout.

5. Analyze and Iterate

Monitor metrics, check for SRM, and analyze segment-level effects. If successful, consider gradual rollout; if not, diagnose and iterate.

Key Points to Mention

  • Two-stage ranking architecture (candidate generation + ranking) to balance latency and relevance.
  • Model optimization techniques: quantization, pruning, distillation, and hardware acceleration (e.g., GPU/TPU).
  • Caching strategies for features and results, and precomputation of embeddings.
  • A/B testing best practices: randomization unit, power analysis, guardrail metrics, and avoiding peeking.
  • Handling network effects and interference in marketplace settings (e.g., Airbnb).
  • Long-term holdout and measuring long-term effects to avoid short-term gains at the expense of user experience.

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