Bourazas, K., "Bayesian binary classification under label uncertainty with network-informed Gaussian Processes"

Title: ""Bayesian binary classification under label uncertainty with network-informed Gaussian Processes""

Speaker: Postdoctoral Researcher Konstantinos Bourazas, Department of Economics, Athens University of Economics and Business.

Host:  Assistant Professor Alexopoulos Angelos, Department of Economics, Athens University of Economics and Business

Time: 15.30 -17.00

Room:  76, Patission Str., Antoniadou Wing, 3rd floor, Room A36

Abstract: We study Bayesian binary classification under one-sided (positive-only) label noise—known as positive–unlabeled (PU) learning—with both covariate and multi-layer network data. We combine these modalities through a Product-of-Experts (PoE) model, where each expert handles one data source, the latent logits combine additively, and the likelihood accounts for missed positives via a false-negative rate η. Our main contributions are: (i) identifiability under mild calibration conditions; (ii) posterior contraction of the latent function at the near-minimax nonparametric rate; and (iii) parametric $n^{-1/2}$ contraction for η. We show the PoE model is distributionally equivalent to a single Gaussian process with composite covariance, and that expert agreement tightens finite-sample bounds. We develop a bespoken Markov chain Monte Carlo algorithm to estimate under the Bayesian paradigm this PU–PoE setting, proving scalability as the number of experts grows. Empirically, our method outperforms strong PU baselines and proves useful in real fuel-tax fraud detection data from the Greek economy, providing calibrated uncertainty for decision support.