Design “Trending” for Articles: Definition, Measurement, and Evaluation

Context

You are building a "Trending" ranking for an articles surface (e.g., home feed, explore). Define what “trending” means, what data you need, and how you would score, evaluate, and safely launch it.

Tasks

1. Definition
   • Define “trending” for articles and how it differs from “popular” or “top”.
2. Data Requirements
   • Specify the raw event schema you require (e.g., article_id, user_id, timestamp, action_type, dwell_time, referrer, device, locale).
3. Scoring
   • Propose a scoring formula that combines signals (impressions, clicks, dwell time, reactions, reshares) with:
     • Recency decay
     • De-duplication (e.g., per user/session)
     • Spam/bot controls
4. Time and Normalization
   • Explain time windows (e.g., 5m, 1h, 24h) and how to baseline-normalize (e.g., by author follower count, historical traffic).
   • Describe category/locale segmentation and cold-start handling for new items.
5. Evaluation and Launch
   • Outline offline evaluation (e.g., precision/recall against editorial ground truth, calibration).
   • Outline online A/B tests, success metrics (CTR, dwell time uplift, save/share rates), and guardrails (session depth, bounce rate, creator fairness).