A1182
Title: Gaussian process regression with transfer learning under distribution shift: A joint probability neural kernel approach
Authors: Ruiwen Cao - University of Leicester (United Kingdom) [presenting]
Bo Wang - University of Leicester (United Kingdom)
Neslihan Suzen - University of Leicester (United Kingdom)
Abstract: Gaussian Process Regression (GPR) is a powerful statistical modeling method, but it typically relies on abundant observational data, which can be both costly to collect in practice. In scenarios where samples for the target task are scarce, GPR with transfer learning (TL) provides a promising solution by leveraging knowledge from the source task to enhance target-task predictions. However, existing GPR with TL methods based on adaptive transfer kernels have several limitations, including optimizing the conditional probability while ignoring the source domain distribution, lacking a correction mechanism to address shifts in feature distributions, and high computational complexity on large-scale datasets. To address these issues, an efficient GPR with TL model, namely TLNKNJP, is proposed with three advantages: (1) joint probability likelihood maximization instead of conditional probability, which improves target-domain prediction while simultaneously preserving accurate inference capability on the source domain; (2) multi-layer Neural Kernel Network (NKN) with Maximum Mean Discrepancy, enabling the model to better adapt to nonlinear relationships and distribution shifts; (3) a novel Wasserstein-based data distillation, reducing computational time for NKN training while preserving source-domain distribution characteristics. Experimental results demonstrate that TLNKNJP significantly outperforms existing methods in terms of prediction accuracy and computational efficiency.
