Cost-Sensitive Support Vector Ranking for Information Retrieval

Abstract

In recent years, the algorithms of learning to rank have been proposed by researchers. However, in information retrieval, instances of ranks are imbalanced. After the instances of ranks are composed to pairs, the pairs of ranks are imbalanced too. In this paper, a cost-sensitive risk minimum model of pairwise learning to rank imbalanced data sets is proposed. Following this model, the algorithm of cost-sensitive supported vector learning to rank is investigated. In experiment, the standard Ranking SVM is used as baseline. The document retrieval data set is used in experiment. The experiment results show that the performance of cost-sensitive support vector learning to rank is better than Ranking SVM on two rank imbalanced data sets.