Airbnb · Primly Community

Airbnb machine learning engineer interview: recsys focus, what they actually asked

ml_mike · 5 replies

went through the airbnb MLE loop about five weeks ago. sharing raw notes.

airbnb is heavily recsys/ranking. if you're coming from pure NLP or CV, you're not disqualified but you need to speak the ranking language. think: how do you balance exploration vs exploitation in a recommendation setting, offline vs online evaluation metrics, feature pipelines for user-item interaction data.

My loop structure:

Coding (2 rounds): one standard coding round, one ML coding round. the standard round was a medium-difficulty graph problem. the ML coding round was more interesting: implement a simplified version of a nearest-neighbor search over embeddings, evaluate a few distance functions, and then discuss how you'd scale this. no sklearn, write the math yourself. they gave me numpy.

ML system design (1 round): i got 'design a ranking system for airbnb search results'. this is as on-brand as it gets. i started with clarifying questions: what's the primary metric (bookings? clicks? satisfaction scores?), what's the latency constraint, what signals do we have. they wanted to see the full pipeline: candidate retrieval, feature engineering, model choice (they asked me to compare gradient boosted trees vs a two-tower neural network for retrieval), offline eval with holdout sets, online A/B test design, and monitoring for model drift. we spent 20 minutes just on the monitoring piece. they care a lot about how you'd detect that your model is degrading in production.

ML depth / research interview: this felt more like a PhD-style discussion. they gave me a paper-adjacent problem about handling position bias in ranking data (users click on top results more regardless of quality). we talked about inverse propensity scoring, unbiased learning-to-rank methods, and logged bandit feedback. if you don't know what position bias is, add it to your prep list.

Behavioral: same format as every other role. "for 'be a host'..." etc.

comp note: the L5 MLE offer i received was base around $200-220k SF range plus RSUs. i framed it as ballpark "a friend got."

prep resources that actually helped: the Airbnb engineering blog (they've posted about their ranking and search systems), the RecSys conference papers on learning-to-rank, and honestly just being able to talk concretely about tradeoffs rather than naming-dropping buzzwords.

5 replies

ds_dmitri

the position bias question is genuinely hard and specific. can you share more about how deep you needed to go on IPS? like did they want the math of the estimator or just the concept?

ml_mike

they wanted the conceptual framing first (what problem does IPS solve, why does naive training on click data create a biased model) and then asked me to sketch the reweighting formula. i could do it at a high level, not perfectly, and they were fine with that. they pushed harder on 'what are the failure modes of IPS in practice' which is: high variance when propensities are very small (rare positions), and you need a large enough dataset. i knew enough to have a real conversation, not memorize a textbook derivation.

growth_gabe

the 'what metric are you optimizing' clarifying question is so important in ML design rounds and people skip it. of course if your primary metric is bookings you design a very different model than if it's listing-quality satisfaction scores.

numbers_only

$200-220k base L5 MLE SF 2026 checks out. RSU refreshers at airbnb have been reasonable post-IPO, heard 4-year vest with 1-year cliff. total comp probably $350-420k depending on stock price at time of grant and grant size.

infra_ines

the model monitoring question is one of those things that separates people who've actually run a prod ML system from people who've only trained models in notebooks. glad to hear they're asking it seriously.