Semi-supervised ranking via ranking regularization

We propose a learning algorithm for ranking functions in the semi-supervised setting, with a training set consisting of a small set of ranked items, and a large set of unranked items. We articulate a regularization principle that exploits unranked data by favoring similar items having similar ranks. This principle is implemented as a semi-supervised extension of LambdaRank. We apply this learning algorithm to a relevance-feedback scenario in information retrieval, and show that semisupervised learning gives improved rankings.