Inherit and Fix a Broken Recommendation System

You join a video platform's recommendations team. The system has been in production for 3 years. The original goal was "keep users engaged." The current system is heavily optimized for watch-time-based proxies. Engineering criticism has documented that the system tends to surface sensational and emotionally loaded content — content that keeps users watching but that users report as draining rather than satisfying.

You are not asked to retrain the model. You are asked to redesign the framing before any model changes happen. Walk through: (1) the business objective you would restate, (2) the proxy labels you would change or supplement, (3) one policy-level constraint or guardrail you would add to the decision layer.

Focus on framing decisions, not model architecture. The crux is identifying what was wrong with the original framing and specifying what "better" looks like in concrete terms.

Follow-up ladder

1. Rung 1: The VP of Product says the new objective ("healthy long-term engagement") is too hard to measure. How do you respond? What intermediate metrics would you propose that are both measurable and directionally correct?
2. Rung 2: You propose adding a user satisfaction signal to the proxy set. The data science team says: "Surveys are sparse and biased — only users who are unhappy fill them out." How do you handle this objection?
3. Rung 3: Six months after the reframe, watch time drops 8% but survey satisfaction scores improve 15%. The CEO wants to know if this was the right call. How do you make the case?