MLSD — Ads Ranking

Design the Meta ads-ranking system. Feature engineering, model architecture, and calibration are the three deep-dive zones reported. Requirements Functional: given a user impression slot, select and rank ads to show; optimize for advertiser value (bid × pCTR × pCVR) while respecting user experience. Pipeline: candidate retrieval (per advertiser targeting) → light pre-rank → heavy ranker (multi-task: pCTR, pCVR, pDwell) → auction (eCPM = bid × predicted value) → re-rank with diversity / frequency cap. Feature engineering: user (demographics, recent activity), ad (creative, advertiser, vertical), context (slot, time, device), cross features, sequence features. Model architecture: DCN-V2 / DLRM for ranker; embedding-based two-tower for retrieval; MMoE for multi-task heads. Calibration: critical for ads — predicted probabilities feed the bid auction. Platt scaling / isotonic regression on a held-out set; per-segment calibration when traffic mix shifts. Metrics: revenue / 1k impressions, advertiser ROI, CTR, conversion lift; long-term user retention guardrail. Notes Multiple candidates report MLSD interviewer focus on feature engineering + calibration over architecture novelty. Calibration is the differentiator vs Reels recommendation — be ready to defend why a well-calibrated model matters for the bid mechanism. One MLE loop spent the entire round on data collection — how training data is sourced, logged, and labeled — rather than model architecture. Be ready to go deep on the data pipeline feeding the ranker, not just the model. Preparation Memorize the eCPM = bid × P(click) × P(conv | click) decomposition. Prep a 3-sentence calibration pitch: why it matters, how to measure (reliability diagram, ECE), how to fix (Platt / isotonic). Pre-write the feature taxonomy (user × ad × context × cross × sequence) so you can list them in 60 seconds.