A feature selection method based on minimum redundancy maximum relevance for learning to rank

Learning to rank has considered as a promising approach for ranking in information retrieval. In recent years feature selection for learning to rank introduced as a crucial issue. Reducing the feature set by removing irrelevant and redundant features can improve the prediction performance. In this paper we address the problem of filter feature selection for ranking. We propose to apply minimum redundancy maximum relevance (mRMR) method that select feature subset based on importance of features and similarity between them. We reweight the component of mRMR to balance between importance and similarity. We apply two methods for measuring the similarity between features and two methods for evaluating importance. Experimental results on two standard datasets from Letor demonstrate that the proposed algorithm 1)outperform two stateof- the-art learning to rank algorithms in term of accuracy, 2) learn a more spars model compared to a feature selection model for ranking.