Improving pairwise learning to rank algorithms for document retrieval

Learning to rank has recently emerged as an attractive technique dedicated to the optimization of results rankings and based on machine learning techniques. Several search engine applications are using this technique to train their ranking model. Theoretically, the problem of ranking is to predict an ordering over a set of pairs query-document. However, the enormously large size of the Web documents makes it generally impossible for the users to find their desired information by surfing the Web. As a consequence, effective and efficient information retrieval (IR) has become more important and also IR system (IRS) has become indispensable tool that can allow the user to access only to the information he deems relevant. Currently, we have proposed two pairwise learning to rank algorithms which combine active and semi-supervised learning to reduce the labeling effort for DR. In this paper, we propose to improve these semi-supervised active algorithms by using a list of pairs in selecting unlabeled data. We showed through different ranking measures that the algorithms proposed yielded into competitive results compared to SAL2R and ASSL2R on collections from the standard benchmark Letor.