An Alternative Cross Entropy Loss for Learning-to-Rank

TL;DR

This paper introduces a new theoretically grounded cross entropy loss for learning-to-rank that is convex, consistent with NDCG, and empirically outperforms existing methods on benchmark datasets.

Contribution

A novel convex, NDCG-consistent cross entropy loss function for learning-to-rank, bridging theory and practice in information retrieval.

Findings

Outperforms existing algorithms in ranking quality.

Demonstrates robustness across benchmark datasets.

Provides a theoretically sound loss function for ranking.

Abstract

Listwise learning-to-rank methods form a powerful class of ranking algorithms that are widely adopted in applications such as information retrieval. These algorithms learn to rank a set of items by optimizing a loss that is a function of the entire set -- as a surrogate to a typically non-differentiable ranking metric. Despite their empirical success, existing listwise methods are based on heuristics and remain theoretically ill-understood. In particular, none of the empirically successful loss functions are related to ranking metrics. In this work, we propose a cross entropy-based learning-to-rank loss function that is theoretically sound, is a convex bound on NDCG -- a popular ranking metric -- and is consistent with NDCG under learning scenarios common in information retrieval. Furthermore, empirical evaluation of an implementation of the proposed method with gradient boosting machines on benchmark learning-to-rank datasets demonstrates the superiority of our proposed formulation over existing algorithms in quality and robustness.