Robust Model Selection for Discovery of Latent Mechanistic Processes

Jiawei Li, Nguyen Nguyen, Meng Lai, Ioannis Ch. Paschalidis, Jonathan H. HugginsarXiv 2026 · DOI · arXiv

Abstract: When learning interpretable latent structures using model-based approaches, even small deviations from modeling assumptions can lead to inferential results that are not mechanistically meaningful. In this work, we consider latent structures that consist of $K_o$ mechanistic processes, where $K_o$ is unknown. When the model is misspecified, likelihood-based model selection methods can substantially overestimate $K_o$ while more robust nonparametric methods can be overly conservative. Hence, there is a need for approaches that combine the sensitivity of likelihood-based methods with the robustness of nonparametric ones. We formalize this objective in terms of a robust model selection consistency property, which is based on a component-level discrepancy measure that captures the mechanistic structure of the model. We then propose the accumulated cutoff discrepancy criterion (ACDC), which leverages plug-in estimates of component-level discrepancies. To apply ACDC, we develop mechanistically meaningful component-level discrepancies for a general class of latent variable models that includes unsupervised and supervised variants of probabilistic matrix factorization and mixture modeling. We show that ACDC is robustly consistent when applied to unsupervised matrix factorization and mixture models. Numerical results demonstrate that in practice our approach reliably identifies a mechanistically meaningful number of latent processes in numerous illustrative applications, outperforming existing methods.

Introduction

In scientific discovery, latent variable models are essential for identifying hidden mechanisms, such as distinct cell types in biological samples or mineral compositions in satellite imagery. A fundamental challenge in these applications is determining the correct number of latent processes ($K$). While standard tools like the Bayesian Information Criterion (BIC) are effective when the model perfectly matches the data, they often fail in the presence of model misspecification. Real-world data rarely follows idealized distributions (e.g., perfect Gaussians); even slight skewness can cause BIC to "overfit," suggesting spurious clusters to compensate for the mathematical mismatch. Conversely, nonparametric methods often lack the sensitivity to detect subtle but genuine signals. This work introduces the Accumulated Cutoff Discrepancy Criterion (ACDC), a framework that balances sensitivity and robustness to identify mechanistically meaningful latent structures.

The ACDC Framework

ACDC shifts the focus from overall predictive fit to component-level stability. The methodology follows a structured approach:

  1. Component-Level Discrepancy: Instead of evaluating the global model fit, ACDC assesses how well each individual discovered component aligns with the model's assumptions using measures such as KL divergence or Wasserstein distance.
  2. The Cutoff Threshold ($\rho$): A user-defined threshold represents the maximum "acceptable" model misspecification. Components with discrepancies below this threshold are considered valid matches to the assumed mechanism.
  3. Accumulation and Selection: ACDC calculates the accumulated discrepancy $R^\rho = \sum \max(0, \text{discrepancy} - \rho)$. The optimal number of processes is the smallest $K$ that minimizes this accumulation.

Geometric Intuition and Robust Consistency

The theoretical foundation of ACDC rests on "robust model selection consistency." As illustrated in Figure 1, we consider nested model manifolds $\mathcal{M}_1 \subset \mathcal{M}_2 \subset \dots \subset \mathcal{M}_{K_o}$. In misspecified settings, the true data components (represented by points $P$) may lie outside these manifolds. ACDC identifies the smallest $K$ where the components are "close enough" to the manifold, effectively ignoring noise induced by model mismatch.

Geometric representation of model misspecification and component-level discrepancy across nested model manifolds.

Experimental Results

The authors validated ACDC across several high-stakes scientific domains, demonstrating its superiority over traditional methods:

  • Single-Cell RNA Sequencing: In tests using the Tabula Muris dataset (~100,000 cells), ACDC achieved a Mean Absolute Error (MAE) of 2.40 in estimating cell types. This significantly outperformed industry standards like Seurat (MAE = 7.21) and SC3 (MAE = 31.23), which both tended to drastically overestimate the number of clusters.
  • Cancer Genomics: When applied to simulated breast cancer data for mutational signature discovery, BIC failed by selecting the maximum possible number of signatures. ACDC correctly identified 7–8 signatures corresponding to known biological processes.
  • Hyperspectral Imaging: In urban material identification, ACDC accurately identified the 5 primary materials (e.g., grass, asphalt), whereas BIC overfitted (suggesting 7) and other methods underfitted (suggesting 3).

Discussion and Conclusion

ACDC prioritizes explanation over prediction. In scientific contexts, a model that minimizes log-loss by adding dozens of non-existent clusters is less valuable than one that identifies the true underlying entities. By providing a formal proof of robust consistency and maintaining computational efficiency—avoiding the expensive simulations required by "coarsened posterior" methods—ACDC offers a practical solution for large-scale genomic and physical data analysis.

Future work will focus on refining the stability-based heuristics for threshold selection and extending the ACDC framework to non-linear model classes, such as Variational Autoencoders (VAEs).

Summary updated

Abstract

When learning interpretable latent structures using model-based approaches, even small deviations from modeling assumptions can lead to inferential results that are not mechanistically meaningful. In this work, we consider latent structures that consist of $K_o$ mechanistic processes, where $K_o$ is unknown. When the model is misspecified, likelihood-based model selection methods can substantially overestimate $K_o$ while more robust nonparametric methods can be overly conservative. Hence, there is a need for approaches that combine the sensitivity of likelihood-based methods with the robustness of nonparametric ones. We formalize this objective in terms of a robust model selection consistency property, which is based on a component-level discrepancy measure that captures the mechanistic structure of the model. We then propose the accumulated cutoff discrepancy criterion (ACDC), which leverages plug-in estimates of component-level discrepancies. To apply ACDC, we develop mechanistically meaningful component-level discrepancies for a general class of latent variable models that includes unsupervised and supervised variants of probabilistic matrix factorization and mixture modeling. We show that ACDC is robustly consistent when applied to unsupervised matrix factorization and mixture models. Numerical results demonstrate that in practice our approach reliably identifies a mechanistically meaningful number of latent processes in numerous illustrative applications, outperforming existing methods.
BibTeX
@misc{li2026robust,
  author = {Jiawei Li and Nguyen Nguyen and Meng Lai and Ioannis Ch. Paschalidis and Jonathan H. Huggins},
  title = {{Robust Model Selection for Discovery of Latent Mechanistic Processes}},
  year = {2026},
  eprint = {2602.22062},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2602.22062},
}
Figure