Snap Ads Ranking

Design an ads ranking system with feature construction, multi-task objectives, delayed labels, calibration, and business constraints. Requirements Design an ads ranking system for a feed or search surface. Functional requirements: Rank candidate ads for a user request using relevance, predicted engagement, advertiser objective, and policy / quality constraints. Build features from user history, ad creative, advertiser, context, and recent behavior. Train models for multiple engagement levels such as click, conversion, hide, long view, or downstream value. Handle delayed and sparse labels, especially conversions. Provide online metrics, offline metrics, and debugging tools for regressions. Scale / constraints: Ranking happens under a tight online latency budget. Labels are biased by position, auction exposure, targeting, and delayed conversion windows. Business objectives may combine user experience, advertiser ROI, revenue, and marketplace health. Design decisions: Retrieval / candidate generation vs final ranking. Single-task CTR model vs multi-task model for several engagement levels. Weighted logistic regression or deep ranking model for different event values. Calibration and auction compatibility. Exploration, cold-start ads, and advertiser budget pacing. Notes A strong answer starts with candidate ads, feature hydration, ranker scoring, filtering, auction / pacing logic, and logging. Multi-task learning is useful when the system optimizes several signals with different sparsity and delay patterns. Weighted logistic regression is a reasonable baseline when engagement types have different business weights; deep models can add cross features and embeddings once the baseline is understood. Delayed labels are central. Use attribution windows, delayed-label correction, negative-label maturity, and online proxy metrics so the model does not learn that every not-yet-converted impression is negative. Calibration matters because downstream auction logic needs probabilities or expected values that mean what they say. Preparation Prepare a diagram with candidate generation, feature store, ranker, auction / pacing, logger, training, and evaluation. Drill features: user interests, ad text / image embedding, advertiser quality, context, frequency caps, and recent negative feedback. Practice explaining delayed labels, calibration, position bias, and multi-task loss weighting. Keep one simple baseline ready before proposing a complex model.