Integrating expert and novice evaluations for augmenting ordinal regression models
Information Fusion - 2019
Abstract
We consider a predictive modelling problem, where the goal is to predict the absolute evaluation of an object on an ordinal scale, traditionally known as an ordinal regression problem. We present a framework that is capable of learning such a model while combining different types of information: absolute evaluations by experts and relative evaluations by novices. We propose and solve a linearly constrained convex optimization problem that takes both types of information into account, and is capable of attributing an ordinal label to a new object based on its features. We do this by relying on principles from machine learning and optimization theory, combined with ideas from information fusion. Experimental results demonstrate the enhanced performance of ordinal regression models when incorporating relative evaluations in the form of rankings.