Learning to rank relational objects based on the listwise approach

Abstract

In recent years machine learning technologies have been applied to ranking, and a new research branch named “learning to rank” has emerged. Three types of learning-to-rank methods - pointwise, pairwise and listwise approaches - have been proposed. This paper is concerned with listwise approach. Currently structural support vector machine(SVM) and linear neural network have been utilized in listwise approach, but these methods only consider the content relevance of an object with respect to queries, they all ignore the relationships between objects. In this paper we study how to use relationships between objects to improve the performance of a ranking model. A novel ranking function is proposed, which combines the content relevance of documents with respect to queries and relation information between documents. Two types of loss functions are constructed as the targets for optimization. Then we utilize neural network and gradient descent algorithm as model and training algorithm to build ranking model. In the experiments, we compare the proposed methods with two conventional listwise approaches. Experimental results on OHSUMED dataset show that the proposed methods outperform the conventional methods.