Model-Agnostic Sample Reweighting for Reliable Strength Behavior Prediction of Coarse-Grained Soils Under Distribution Shifts

Ying Zhang, Ziying An, Xuedong Zhang, Liping Cao, Jingzhe Jiao, Shaohao Tan

Keywords: Coarse-grained soils, Mechanical property analysis, Neural network, Invariant learning, Sample reweighting, Out-of-Distribution Generalization

TL;DR: We propose Model-Agnostic Sample Reweighting (MASR), a bilevel-optimized framework that integrates sample reweighting and invariant learning to enhance OOD generalization and model interpretability in coarse-grained soil strength prediction.

Abstract: Coarse-grained soils are widely employed in geotechnical engineering due to their favorable compaction and load-bearing properties. However, accurately modeling their strength behavior, typically characterized by the deviatoric stress–axial strain $(q - \varepsilon_a)$ relationship, remains challenging under real-world conditions, particularly when training and testing data exhibit distribution shifts. Existing deep learning approaches often rely on the independent and identically distributed (i.i.d.) assumption, limiting their robustness and generalization to out-of-distribution (OOD) scenarios. To address this limitation, we propose Model-Agnostic Sample Reweighting (MASR), a novel framework that integrates sample reweighting and invariant learning via bilevel optimization. In this approach, the inner loop optimizes the model on weighted samples, while the outer loop adjusts these weights to enhance generalization under OOD conditions. This iterative mechanism enables the model to progressively identify invariant and causally relevant features while suppressing spurious correlations. Empirical evaluations on real-world coarse-grained soil datasets demonstrate that MASR outperforms baseline models in predictive accuracy under significant distribution shifts.

Primary Area: causal reasoning

Submission Number: 7880