The discounted cumulative margin penalty: Rank-learning with a list-wise loss and pair-wise margins

Abstract

In recent years, the fields of rank-learning and information retrieval have received substantial attention. Algorithms developed within these domains have shown promising results in a variety of problem spaces, especially in document retrieval and web search. In this paper, a new rank-learning algorithm is proposed that combines list-wise loss measurements with pair-wise margins. The list-wise loss term is inspired by the Normalized Discounted Cumulative Gain (NDCG) metric, and the resulting objective function is solvable with gradient-free optimization techniques. Experiments using the LETOR 3.0 and 4.0 collections demonstrate that the ranking performance achieved by an algorithm using this loss measure is competitive with reported baselines.