Learning-to-Rank: Features, Models, and Position Bias

~45 min read

The three learning-to-rank paradigms, feature engineering for search, why interaction features dominate, and how to correct position bias when training on click log data.

Learning-to-Rank: Features, Models, and Position Bias is a Pro topic

Sign in, then upgrade to Pro or Power to unlock this topic and the full AI Engineering curriculum.

Key points

  • LambdaRank (pairwise LTR with NDCG-weighted gradients) is the standard production choice: more principled than pointwise, more tractable than listwise
  • Query-document interaction features (BM25 score, semantic similarity, historical CTR for this query-document pair) consistently dominate document quality features — always build these first
  • Position bias is systematic: clicks at position 1 are 5-10× more frequent than position 5 for the same document. IPS correction re-weights training examples to de-bias the model