9.7.3
ListMLE, ListNet, LambdaRank and RankNet
Compare the principal listwise and rank-based loss functions described in the source.
What is it?
ListMLE, ListNet, LambdaRank and RankNet are ranking loss functions with different approaches to learning from ranked responses.
Key points
ListMLE: directly maximises the Plackett-Luce likelihood of the ground-truth ranking.
ListNet: minimises KL divergence between the model top-1 probability distribution and the ground-truth distribution.
LambdaRank: weights pairwise gradients by changes in ranking metrics such as NDCG; useful when top-ranking quality matters more.
RankNet: sums pairwise cross-entropy across all pairs, equivalent to Bradley-Terry on the pairs extracted from the ranking.
Choosing a Ranking Loss
If the goal is to learn the complete ground-truth ordering, ListMLE directly optimises the Plackett-Luce likelihood. If the top-ranked response matters most, LambdaRank can weight gradient updates according to their effect on a ranking metric such as NDCG.
Real-world application
Choose the loss according to the training signal and ranking objective: complete ranking likelihood, top-item probability, metric-sensitive ranking or pairwise comparisons.
Key takeaway
Different ranking losses optimise different aspects of the ranking signal.