Learning to rank with voted multiple hyperplanes for documents retrieval

The central problem for many applications in Information Retrieval is ranking. Learning to rank has been considered as a promising approach for addressing the issue. In this paper, we focus on applying learning to rank to document retrieval, particularly the approach of using multiple hyperplanes to perform the task. Ranking SVM (RSVM) is a typical method of learning to rank. We point out that although RSVM is advantageous, it still has shortcomings. RSVM employs a single hyperplane in the feature space as the model for ranking, which is too simple to tackle complex ranking problems. In this paper, we look at an alternative approach to RSVM, which we call "Multiple Vote Ranker" (MVR), and make comparisons between the two approaches. MVR employs several base rankers and uses the vote strategy for final ranking. We study the performance of the two methods with respect to several evaluation criteria, and the experimental results on the OHSUMED dataset show that MVR outperforms RSVM, both in terms of quality of results and in terms of efficiency.