eBay · Primly Community

eBay machine learning engineer interview: recsys depth and practical ML

ml_mike · 5 replies

Did the eBay MLE loop about two months ago. Sharing because there's a lot of generic 'ML interview tips' out there and not much specific to eBay, which has a genuinely interesting ML problem space (search ranking, recommendations, fraud detection, price prediction).

Process was: recruiter screen, ML technical phone screen, four-round onsite (virtual). Total time about 7 weeks including a slow debrief period.

Phone screen was 45 minutes. One coding problem (medium LeetCode, arrays) and one ML question: 'walk me through how you'd build a click-through rate model for eBay search results.' Not super deep, more to see if you can talk through a problem coherently. Feature engineering, training pipeline, offline evaluation, A/B test considerations.

Onsite round 1 was coding. Two problems, DSA focus, similar to what you'd see in a general SWE loop. Nothing ML-specific. I got a graph traversal and a dynamic programming problem. Hard-medium difficulty.

Onsite round 2 was ML system design. The prompt: design eBay's recommendation system for 'similar items on the listing page.' I talked through candidate generation (ANN retrieval over embeddings), ranking (LambdaRank or a gradient boosted ranker), feature store, serving latency requirements, and online/offline evaluation. The interviewer pushed hard on: how do you handle popularity bias, how do you evaluate in the absence of a clean ground truth, cold start for new listings. This was the most technical and most interesting round.

Onsite round 3 was ML theory and concepts. Not a vibe check, genuinely deep. Questions: explain gradient boosting and how you'd tune it, what's the difference between L1 and L2 regularization in practice (not just the formula), how do you handle class imbalance in a fraud detection setting, what's feature leakage and how have you caught it. I've been doing this 8 years and a couple of these made me think.

Onsite round 4 was behavioral. Standard STAR. They asked specifically about a time I disagreed with a product decision on what data to use. That question seemed like a real scenario they'd dealt with.

Compensation: my offer was for a senior MLE role, San Jose, $230K base, total package around $380K with RSUs at the 4-year grant. I had one competing offer at a smaller company at higher base but lower total. Accepted eBay for the problem domain.

5 replies

ds_dmitri

The popularity bias question in recsys is a classic but a lot of people don't have a crisp answer ready. Exposure debiasing, counterfactual evaluation, or just downsampling popular items in training. Did they push on a specific method or were they open to different approaches?

ml_mike

Open to different approaches. They seemed more interested in whether you knew the bias existed and could reason through mitigations vs. expecting one specific answer. I talked about inverse propensity scoring and they seemed satisfied.

infra_ines

Seven weeks is long. Did the slow debrief phase cause any issues with other offers you had in flight?

quietquit_quincy

The 'disagree with a product decision on data' behavioral is interesting. Most MLE roles don't explicitly probe for that. Seems like they've had issues with eng vs. product alignment on data choices.

returner_ren

That total comp is pretty solid for eBay actually. I had them in my research as a lower-paying employer but maybe that's changed or it's leveling specific.