One score is rarely enough
Optimizing only for clicks invites clickbait; optimizing only for watch time may bury short useful items. Real products care about several goals at once, so multi objective ranking combines multiple predicted outcomes into a single ordering.
The objectives in play
- Engagement: click, watch, or purchase probability.
- Satisfaction: dwell time, completion, or explicit ratings.
- Long term value: retention and healthy use, not just the next tap.
- Constraints: fairness, diversity, and creator ecosystem health.
Combining predictions
- A common approach predicts each objective separately, then forms a weighted combination as the final score.
- Weights are tuned by experiment to hit a target balance, often using a small set of guardrail metrics.
- More advanced setups frame it as constrained optimization, maximizing one goal subject to floors on others.
The weighting challenge
Weights encode product values and are not learned from clicks alone, since clicks do not reveal long term satisfaction. Teams set them deliberately and validate online.
Watching for conflict
Objectives can pull in opposite directions. Tracking each metric separately during launches reveals whether a gain in one is quietly hurting another.
Key idea
Multi objective ranking blends several predicted outcomes such as engagement, satisfaction, and long term value into one score, with weights set to encode product values and validated online.