Semi-supervised Ranking with Nonsmooth Cost Functions

We propose a semi-supervised learning to rank algorithm. The input is a training set consisting of a small set of preferences or labels, and a large set of unlabeled items. The method directly optimizes popular ranking metrics, such as NDCG, mean average precision, or AUC. It is implemented as a semi-supervised extension of LambdaRank, and achieves near-linear time performance in the number of labeled and unlabeled items. Experimental results show that semi-supervised ranking keeps benefitting from unlabeled data even when thousands of preferences are available.