A probabilistic framework for the design of instance-based supervised ranking algorithms in an ordinal setting |
| |
Authors: | S Lievens B De Baets K Cao-Van |
| |
Institution: | (1) Department of Applied Mathematics and Computer Science, Ghent University, Krijgslaan 281 S9, 9000 Ghent, Belgium;(2) Department of Applied Mathematics, Biometrics and Process Control, Ghent University, Coupure links 653, 9000 Ghent, Belgium |
| |
Abstract: | In this article, we present a probabilistic framework which serves as the base from which instance-based algorithms for solving
the supervised ranking problem may be derived. This framework constitutes a simple and novel approach to the supervised ranking
problem, and we give a number of typical examples of how this derivation can be achieved.
In this general framework, we pursue a cumulative and stochastic approach, relying heavily upon the concept of stochastic
dominance. We show how the median can be used to extract, in a consistent way, a single (classification) label from a returned
cumulative probability distribution function. We emphasize that all operations used are mathematically sound, i.e. they only
make use of ordinal properties.
Mostly, when confronted with the problem of learning a ranking, the training data is not monotone in itself, and some cleansing
operation is performed on it to remove these ‘inconsistent’ examples. Our framework, however, deals with these occurrences
of ‘reversed preference’ in a non-invasive way. On the contrary, it even allows to incorporate information gained from the
occurrence of these reversed preferences. This is exactly what happens in the second realization of the main theorem. |
| |
Keywords: | Instance-based learning Monotone classification Supervised ranking Stochastic dominance |
本文献已被 SpringerLink 等数据库收录! |
|