Yu-Ting Liu (Microsoft Research Asia and Beijing Jiatong University)
Tie-Yan Liu (Microsoft Research Asia)
Tao Qin (Tsinghua University)
Zhi-Ming Ma (Chinese Academy of Science)
Hang Li (Microsoft Research Asia)
This paper is concerned with rank aggregation, the task of combining results of individual ranking functions in meta-search. Previously, rank aggregation was performed mainly by using unsupervised methods. It is hard for the unsupervised approach to improve ranking performances by leveraging the use of labeled data, when such data is available. We propose employing a supervised learning approach to perform the task, which we refer to as "Supervised Rank Aggregation". We set up a general framework for conducting rank aggregation with supervised learning, in which learning for rank aggregation is formalized as an optimization issue that minimizes disagreements with the labeled ground truth data. As case study, we focus on Markov Chain based rank aggregation in this paper. The optimization problem is not a convex optimization problem for Markov Chain based methods, however, and thus is hard to solve. We transform the optimization problem into semi-definite programming and give proofs on the correctness. Experimental results on meta-searches show that Supervised Rank Aggregation can significantly outperform existing unsupervised methods.
Alberta, Thursday, May 10, 2007, 3:30pm to 5:00pm.