Rankingmeasures And Loss Functions In Learning To Rank

By Wei Chen et al

Read the original document by opening this link in a new tab.

Table of Contents

1 Introduction
2 Related Work
3 Main Results
3.1 Essential Loss: Ranking as a Sequence of Classifications
3.2 Essential Loss: Upper Bound of Measure-Based Ranking Errors
3.3 Essential Loss: Lower Bound of Loss Functions

Summary

Learning to rank has become an important research topic in machine learning. Most methods learn ranking functions by minimizing loss functions, but ranking measures like NDCG and MAP are used to evaluate performance. This work reveals the relationship between ranking measures and loss functions, showing that minimizing certain loss functions leads to maximization of ranking measures. The essential loss is introduced as a key concept, which is proven to be an upper bound of measure-based ranking errors. Experimental results validate the theoretical analysis by demonstrating improved ranking performance. The paper discusses different approaches in learning to rank, including pointwise, pairwise, and listwise methods. It also explores widely-used loss functions and ranking measures in information retrieval. Theoretical results are presented on the relationship between loss functions and ranking measures, showing upper bounds for certain loss functions. The study provides insights into optimizing ranking measures through loss function modifications.