Reproduce “An Alternative Cross Entropy Loss for Learning-to-Rank” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Sebastian Bruch
Sebastian Bruch from Google Research/NIH introduces xe_ndcg, an alternative cross-entropy loss for learning-to-rank, proving it is a convex bound on NDCG and demonstrating its NDCG-consistency under common IR conditions. The proposed loss, when optimized with LightGBM, consistently outperforms state-of-the-art LambdaMART in ranking quality, convergence speed, and robustness to various types of noise on benchmark datasets.
Key insights
- xe_ndcg is proven to be a convex upper bound on negative (translated and log-transformed) mean NDCG, establishing a direct analytical link between the loss and the ranking metric.
- Unlike LambdaMART, whose complexity scales quadratically with list length, xe_ndcg is 2-Lipschitz, indicating constant complexity and better generalization properties.
- The proposed loss is NDCG-consistent under common information retrieval learning scenarios, ensuring its optimal solution aligns with the target ranking metric.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.