Regime-Aware Semi-Supervised Regression via Clustering-Gated Experts

ICLR 2026 Conference Submission15576 Authors

Keywords: semi-supervised regression; mixture of experts; regime discovery; clustering-gated routing; tabular learning; uncertainty estimation; pseudo-label risk control; consistency regularization; conditional shift; industrial analytics

Abstract: We study regime-aware semi-supervised \emph{regression} for tunnel boring machine (TBM) operation modeling under cross-strata nonstationarity and label scarcity. We propose \textbf{CGE}—\emph{Clustering-Gated Experts}—a three-stage framework that (i) discovers latent geological regimes via robust, ensemble clustering; (ii) trains per-regime heterogeneous ensembles with agreement-based pseudo-labeling and consistency regularization; and (iii) routes predictions by a lightweight distance-based soft gate. For risk-aware deployment, we equip all predictors with conformalized quantile regression (CQR) to yield calibrated prediction intervals. On real TBM data with 5–20\% label budgets, \textbf{CGE} surpasses strong semi-supervised baselines, achieving at 10\% labels an average $R^2$ of $\mathbf{0.942}\!\pm\!0.018$ and RMSE of $\mathbf{0.112}\!\pm\!0.015$. With 90\% CQR intervals, it attains near-nominal coverage and the narrowest widths, alongside lower NLL/CRPS. Overall, \textbf{CGE} offers a practical accuracy–uncertainty trade-off for safety-critical TBM decision-making under nonstationary geology.

Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)

Submission Number: 15576