MLSD — Notification Ranking (Multi-source)

Rank notifications from heterogeneous sources (friend activity, ads, system) under a unified objective. Multi-objective ranking + frequency-cap + value-normalization across sources are the deep-dive axes. Requirements Functional: each user has a notification inbox fed by many sources (social, ads, product, system). Decide ordering + send/skip. Inputs per candidate: source type, freshness, predicted CTR, advertiser value (if applicable), user fatigue. Pipeline: source-specific retrievers → unified candidate pool → multi-objective ranker → frequency-cap + diversity post-filter → delivery channel selection. Metrics: per-source CTR, daily-active notifications, unsubscribe / complaint rate. Cold start: new user gets product-default notifications + content-based fallback. Notes Reported emphasis: cross-source value normalization. "How do you compare an ad bid in $ to a friend-event CTR?" Calibrated probability × monetization weight is the standard answer. Frequency cap is often the deciding deep-dive — global per-user cap vs per-source caps vs adaptive cap. Bias and feedback-loop discussion expected: high-CTR notifications crowd out lower-CTR ones forever without exploration. Preparation Memorize the multi-source unified-pool diagram. Prep a 2-minute value-normalization pitch (calibrated probability × value weight + per-source clipping). Drill the frequency-cap deep-dive: token bucket per user × source × type.