My recommender is giving too many popular items

That usually means the model is over-indexing on global popularity and not enough on personal preference, freshness, or diversity.

A few common causes and fixes:

Common causes

• Popularity bias in training data: popular items get more clicks/purchases, so the model learns to favor them.
• Objective function rewards easy wins: ranking loss may prefer items with strong average signal.
• Insufficient personalization features: user/item interactions may not be rich enough.
• Candidate generation too narrow: the retrieval stage may already be mostly popular items.
• Feedback loops: popular items get more exposure, which makes them even more popular.

Ways to reduce it

• Downweight popularity in ranking or scoring.
• Add re-ranking constraints:
  • cap the number of popular items per list
  • diversify by category/brand/price
  • inject novelty or freshness
• Use popularity as a feature, not a dominant signal.
• Debias training data:
  • exposure-aware learning
  • inverse propensity weighting
  • negative sampling that isn’t dominated by popular items
• Improve personalization:
  • user embeddings / sequence models
  • context features
  • session intent
• Broaden retrieval:
  • multiple candidate sources
  • semantic similarity / content-based retrieval
• Tune the evaluation metric:
  • track coverage, long-tail share, novelty, and personalization metrics, not just CTR/NDCG

Practical quick fixes

If you want something fast:

1. Add a popularity penalty in the final score.
2. Re-rank top-k with a diversity/novelty constraint.
3. Mix in a percentage of long-tail items for exploration.
4. Measure exposure distribution to confirm the skew.

If you want, I can help you diagnose this more precisely if you share:

• your model type,
• training objective,
• retrieval/ranking setup,
• and a sample of the top-ranked items.