EGRank: An exponentiated gradient algorithm for sparse learning-to-rank

This paper focuses on the problem of sparse learning-to-rank, where the learned ranking models usually have very few non-zero coefficients. An exponential gradient algorithm is proposed to learn sparse models for learning-to-rank, which can be formulated as a convex optimization problem with the ℓ1 constraint. Our proposed algorithm has a competitive theoretical worst-case performance guarantee, that is, we can obtain an ϵ-accurate solution after O(1ϵ) iterations. An early stopping criterion based on Fenchel duality is proposed to make the algorithm be more efficient in practice. Extensive experiments are conducted on some benchmark datasets. The results demonstrate that a sparse ranking model can significantly improve the accuracy of ranking prediction compared to dense models, and the proposed algorithm shows stable and competitive performance compared to several state-of-the-art baseline algorithms.