Ads Click Prediction

Source: 30-System-Design/Framework & Systems

Prompt

• Predict CTR for candidate ads given user, context, and ad features.

Requirements

• Functional: accurate CTR prediction, calibration, cold start handling.
• Non-functional: millisecond latency, high throughput, fairness.
• Constraints: strict p95 latency and tight budget for feature lookups.

Success Metrics

• Offline: log loss, AUC, calibration error.
• Online: CTR, revenue, conversion rate, advertiser ROI.
• Guardrails: user experience (skip/hide), policy compliance.

Data

• Sources: impressions, clicks, conversions, ad metadata.
• Labeling: click within attribution window, conversion signals.
• Leakage risks: post-click features, delayed conversion signals.

Modeling

• Baselines: logistic regression, GBDT.
• Deep models: wide & deep, DeepFM, DIN.
• Feature crosses: categorical embeddings + interactions.
• Calibration: isotonic or Platt scaling on top of ranker.

Serving

• Retrieval: candidate generation from ads index.
• Scoring: batch + real-time features in feature store.
• Latency budget: strict p95/p99 targets.
• Fallback: cached scores for popular ads/segments.

Evaluation & Monitoring

• Offline eval: time-based splits, leakage checks.
• Online eval: A/B tests with revenue + guardrails.
• Drift/abuse: advertiser gaming, distribution shift.
• Monitoring: calibration drift and feature null rates.

Risks & Tradeoffs

• Calibration vs ranking performance.
• Revenue vs user experience.
• Short-term CTR vs long-term retention.

Notes