Doctor of Philosophy (PhD), Wright State University, 2015, Computer Science and Engineering PhD
In this thesis, we discuss two issues in the learning to rank area, choosing effective objective loss function, constructing effective regresstion trees in the
gradient boosting framework, as well as a third issus, applying learning to rank models into statistcal machine translation.
First, list-wise based learning to rank methods either directly optimize performance measures or optimize surrogate functions of performance measures that have
smaller gaps between optimized losses and performance measures, thus it is generally believed that they should be able to lead to better performance than point-
and pair-wise based learning to rank methods. However, in real-world applications, state-of-the-art practical learning to rank systems, such as MART and
LambdaMART, are not from list-wise based camp. One cause may be that several list-wise based methods work well in the popular but very small LETOR datasets but
fail in real-world datasets that are often used for training practical systems.
We propose a list-wise learning to rank method that is based on a list-wise surrogate function, the Plackett-Luce (PL) model. The PL model has convex loss to
ensure a global optimal guarantee, and is proven to be consistent to certain performance measures such as NDCG score. When we conduct experiments on the PL
model, we observe that it is actually unstable in performance; when the data has rich enough features, it gives very good results, but for data with scarce
features, it fails horribly. For example, when we apply the PL with a linear model on the Microsoft 30K dataset, it gives 7.6 points worse NDCG@1 score than an
average performance of several linear systems. This motivates us to propose our new ranking system, PLRank, that is suitable for any data sets through a mapping
from feature space into tree space to gain more expressive power. PLRank is trained based on the gradient boosting framework, and it is simple to implement. It
has th (open full item for complete abstract)
Committee: Shaojun Wang Ph.D. (Advisor); Keke Chen Ph.D. (Committee Member); Xinhui Zhang Ph.D. (Committee Member); Michael Raymer Ph.D. (Committee Member)
Subjects: Computer Engineering