On Sparse Multi-Task Gaussian Process Priors for Music Preference Learning
In this paper we study pairwise preference learning in a music setting with multitask Gaussian processes, and examine the effect of sparsity in the input space as well as in the actual judgments. To introduce sparsity in the inputs, we extend a classic pairwise likelihood model to support sparse, multi-task Gaussian process priors based on the pseudo-input formulation. Sparsity in the actual pairwise judgments is obtained by a simple greedy sequential design approach. We combine the sequential approach with the pseudo-input preference model and demonstrate various properties of the combined model on synthetic examples. A preliminary simulation shows the performance on a real-world music preference dataset motivating and showing the potential of the sparse Gaussian process formulation for pairwise likelihoods.