A GP classification approach to preference learning

In this abstract I present the problem of learning from pairwise preference judgements as a special case of binary classification. I discuss why kernel classifiers using traditional kernels based on the distance between items cannot be used to address the problem effectively: the preference prediction problem has inherent symmetry properties that these kernels cannot model. I will review a hierarchical Bayesian model for preference learning that by construction respects these symmetry properties and show its equivalence to probit Gaussian process (GP) classification. Motivated by this model I define the preference judgement kernel and show that it is capable of modelling the symmetry of preference judgement.