HMTMO-GP: Hierarchical Multi-Task Multi-Output Gaussian Processes

Yan Chen, Ti-chiun Chang, Kevin Stone

Keywords: Probabilistic Forecasting, Gaussian Processes, Multi-Task Learning, Multi-Output Learning, Hierarchical Modeling, Uncertainty Quantification

TL;DR: HMTMO-GP is a hierarchical Gaussian process framework that models multi-task, multi-output data by combining LMC-style kernel mixing with hierarchical sharing to improve forecasting under structured heterogeneity.

Abstract: Machine learning models often employ a multi-task/multi-output (MTMO) architecture to tackle data scarcity by exploiting structural correlation. In this work, we extend Gaussian processes (GP) for hierarchical MTMO problems, referred to as HMTMO-GP, which leverages GP probabilistic modeling for data-efficient learning and intrinsic uncertainty estimation. Our results on gene expression time series and an antibody solution viscosity dataset demonstrate that HMTMO-GP delivers strong predictive accuracy and well‑calibrated uncertainty estimates compared to existing multi‑task or multi‑output GP models.

Submission Number: 20