Short-Term Scientific Mission

Optimising Portfolios for Multi-Layer Stacking

Main theme: Algorithm Selection and Configuration
Grantee: Nick Kocher, Rheinisch-Westfälische Technische Universität Aachen, Aachen, Germany
Host: Carola Doerr, Sorbonne Université, Paris, France

Start date: 2026-09-08
End date: 2026-09-25
Awarded: 2026-07-30

Description

Modern automated machine learning frameworks such as AutoGluon forgo algorithm configuration in favour of a single, meta-learned portfolio reused across all layers of a multi-layer stacking ensemble. Portfolio algorithms are currently selected only for predictive performance, even though base-layer models are rarely used directly for the final prediction. The STSM will explore multi-objective portfolio selection so that base-layer predictions become more informative for later layers. Stacking is also prone to stacked overfitting, a subtle leakage effect in which predictions from earlier layers carry information into the next layer even for untrained instances. Constrained Bayesian optimisation will be investigated as a way of selecting higher-layer models less susceptible to this leakage.