Learning to rank order

Abstract

The problem of learning to rank order data is a branch of machine learning, and has received a lot of interests in recent years. It has applications in many domains such as information retrieval, recommender systems, decision making, medicine, and financial portfolio construction. This Thesis studies this problem from different perspectives with the aim of advancing the state of the art.

There are four parts in this Thesis. In the first part (Chapter 2) the field of learning to rank order is reviewed, with particular focus on two important subfields, namely ranking creation and ordinal regression.

The second part (Chapters 3–5) focuses on findings in the area of ranking creation. Distance-based ranking creation is introduced in order to support the hypothesis that metric learning is applicable in the area of learning to rank order. Also ranking creation based on pairwise preference function is discussed and the notion of “strength” of preference is introduced. The merit of applying strength (in addition to direction) of preference in the training process is supported by experimental results.

The third part of this Thesis (Chapter 6) focuses on a describing a novel technique for the ordinal regression problem, based on pairwise preferences. The pairwise preference function introduced into the ordinal regression problem also provides information about strength of preference, and not only its direction. Also a novel technique for turning partial rankings (information about pairwise dependencies) into total ranking (ordinal regression category labels) is introduced. The usefulness of the technique is supported by experimental results.

The last part (Chapter 7) provides summary of this Thesis and discusses possible future direction of this research based on constructive neural networks.

Date of Award	Feb 2011
Original language	English
Supervisor	Hui Wang (Supervisor) & William Blackburn (Supervisor)