← Lessons

quiz vs the machine

Silver1120

Machine Learning

The Freshness and Recency

Giving new content a fair shot without flooding the feed.

5 min read · intro · beat Silver to climb

Why time matters

In news, social, and short video, an item's value can decay within hours. A model trained on engagement alone underrates fresh items because they have little history yet, while overrating old viral hits. Handling recency keeps feeds timely and gives new content a chance.

The cold start of new items

  • A brand new item has no clicks, so a pure engagement model scores it low.
  • Without help it never gets shown, so it never earns the data to prove itself, a chicken and egg trap.
  • Freshness boosts and exploration break this loop.

Techniques for freshness

  • Recency features: feed item age into the model so it can learn time sensitive patterns.
  • Time decay: weight older engagement less so stale popularity fades.
  • Freshness boosts: add a controlled bonus to new items in re ranking.
  • Dedicated fresh sources: a retrieval channel just for recent items.

Balancing the tradeoff

Too much freshness floods users with unproven content; too little buries timely items. The right level is tuned by watching long term engagement, not raw click counts on day one.

Key idea

Freshness handling uses recency features, time decay, and controlled boosts to give new content a fair chance and keep feeds timely without flooding users with unproven items.

Check yourself

Answer to earn rating on the learn ladder.

1. Why does a pure engagement model underrate fresh items?

2. What does time decay weighting accomplish?