Keke Chen (Committee Member), Shaojun Wang (Advisor), Xinhui Zhang (Committee Member)
Master of Science (MS)
The main challenge in learning-to-rank for information retrieval is the difficulty to di- rectly optimize ranking measures to automatically construct a ranking model from training data. It is mainly due to the fact that the ranking measures are determined by the order of ranked documents rather than the specific values of ranking model scores, thus they are non-convex, nondifferentiable and discontinuous. To address this issue, listwise approaches have been proposed where loss functions are defined either by exploiting a probabilistic model or by optimizing upper bounds or smoothed approximations of ranking measures. Even though very promising results have been achieved, there is still a mismatch between target cost and optimization cost. In this work, we present a novel learning algorithm that directly optimizes the ranking measures without resorting to any upper bounds or approx- imations. Our approach is essentially an iterative greedy coordinate descent method in optimization. For each iteration, we only update one parameter along one coordinate with all others fixed. Since the ranking measure is a stepwise function of a single parameter, we exploit an exhaustive line search algorithm to locate the interval with the smallest ranking measure along each coordinate. We pick the coordinate that leads to the largest reduction of ranking measure. In order to determine the optimal value of the parameter for the selected coordinate, we construct a probabilistic framework for the permutation, and maximize the likelihood of top-m ranked documents. This iterative procedure is continued until conver- gence. We conduct experiments of five datasets selected from Microsoft LETOR datasets, our experimental results show that the proposed direct rank algorithm outperforms several well-known state-of-the-art ranking algorithms.
Department or Program
Department of Computer Science
Year Degree Awarded
Copyright 2012, all rights reserved. This open access ETD is published by Wright State University and OhioLINK.