← Lessons

quiz vs the machine

Gold1400

Machine Learning

The Re ranking and Diversity

Why the top scored list is not always the best list to show.

5 min read · core · beat Gold to climb

Beyond pointwise scores

Ranking scores each item alone, but a feed is consumed as a set. Showing five near identical items, each scored highly, can feel repetitive and lower overall satisfaction. Re ranking adjusts the final list to account for how items interact.

Goals of re ranking

  • Diversity: avoid stacking near duplicate items so the feed feels varied.
  • Business rules: enforce caps, promotions, or freshness requirements.
  • Slot effects: account for the fact that nearby items influence each other.

Maximal marginal relevance

A classic method, maximal marginal relevance, greedily picks the next item to maximize relevance minus a penalty for similarity to items already chosen. A tuning weight balances pure relevance against variety.

Determinantal point processes

More advanced systems use determinantal point processes, which model a whole set and assign higher probability to sets that are both relevant and mutually dissimilar, capturing diversity globally rather than item by item.

The tradeoff

Too much diversity buries the best items; too little creates monotony. The right balance is found by measuring downstream engagement, not by the raw scores alone.

Key idea

Re ranking treats the feed as a set, trading some pointwise relevance for diversity and rule compliance so the overall experience beats a naive top scored list.

Check yourself

Answer to earn rating on the learn ladder.

1. Why is the top scored list sometimes not the best feed to show?

2. What does maximal marginal relevance do?