An efficient Ranking SVM based on the transitivity of partial order

Abstract

Learning to rank is one of the hottest topics in Information Retrieval (IR) field. Ranking SVM (RSVM) is a typical method of learning to rank. But this approach is time consuming, which decreases its applicability in real-world IR applications, which involves a large amount of computation, because it requires increasing the complexity from n to O(n2). This paper analyzes the characteristics of the partial order on instance pairs. We point out and prove that there is a transitive characteristic in this kind partial order data. An improved loss function is proposed based on the transitivity, which reduces the complexity greatly. Also we give the bound of the complexity of the improved RSVM, from which we can see that the actual complexity is usually near the lower bound in real-world application. Experimental results show that our method, efficient Ranking SVM (eRSVM), out-perform the traditional method RSVM in efficiency greatly without decreasing the ranking accuracy.