LambdaLoss

Stub. Created as a placeholder — expand with vault-sourced content.

A generalized listwise ranking objective framework (Wang et al., Google, 2018) that frames learning-to-rank as a mixture model where the ranked list π is a hidden variable, with the loss defined as the negative log-likelihood of that model.

Its two key results:

1. Existing methods — RankNet, LambdaRank, SoftRank, ListNet — are special configurations of the framework (obtained by choosing the likelihood p(y | s, π) and list distribution p(π | s)).
2. It enables metric-driven loss functions directly tied to the target ranking metric (NDCG), achieving state-of-the-art accuracy.

This positions LambdaLoss as the unifying generalization above the classic pointwise/pairwise/listwise split, and a principled successor to the LambdaMART family.

Related Concepts

• Ranking Objectives · Learning to Rank · LambdaMART · RankNet · NDCG

Related Articles

• Learning to Rank - A Complete Guide to Ranking using Machine Learning — introduces the framework