Unrestricted permutation forces extrapolation: variable importance requires at least one more model, or there is no free variable importance

Hooker, Giles; Mentch, Lucas; Zhou, Siyu

doi:10.1007/s11222-021-10057-z

Cited by 120 publications

(66 citation statements)

References 32 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Since PD plots extrapolate to potentially unobserved areas in the predictor space, they have been critized for being unreliable when predictors are correlated (Apley & Zhu, 2016;Hooker et al, 2021;Molnar, 2019;Molnar et al, 2020;Scholbeck, 2018). Therefore, we will look into the effects of correlated predictors on PD curves further below.…”

Section: Partial Dependence Plotsmentioning

confidence: 99%

Interpretable machine learning for psychological research: Opportunities and pitfalls

Henninger¹,

Debelak²,

Rothacher³

et al. 2022

Preprint

View full text Add to dashboard Cite

In recent years, machine learning methods have become increasingly popular prediction methods in psychology. At the same time, psychological researchers are typically not only interested in making predictions about the dependent variable, but also in learning which predictor variables are relevant, how they influence the dependent variable, and which predictors interact with each other. However, most machine learning methods are not directly interpretable. Interpretation techniques that support researchers in describing how the machine learning technique came to its prediction may be a means to this end. We present a variety of interpretation techniques and illustrate the opportunities they provide for interpreting the results of two widely used black box machine learning methods that serve as our examples: random forests and neural networks. At the same time, we illustrate potential pitfalls and risks of misinterpretation that may occur in certain data settings. We show in which way correlated predictors impact interpretations with regard to the relevance or shape of predictor effects and in which situations interaction effects may or may not be detected. We use simulated didactic examples throughout the article, as well as an empirical data set for illustrating an approach to objectify the interpretation of visualizations. We conclude that, when critically reflected, interpretable machine learning techniques may provide useful tools when describing complex psychological relationships.

show abstract

Section: Partial Dependence Plotsmentioning

confidence: 99%

Interpretable machine learning for psychological research: Opportunities and pitfalls

Henninger¹,

Debelak²,

Rothacher³

et al. 2022

Preprint

View full text Add to dashboard Cite

show abstract

“…, X p )]. As argued for example by Kumar et al (2020) and Hooker et al (2021), care is needed in applying and interpreting feature-wise influence metrics as they may rely on the fitted model extrapolating beyond the sample support.…”

Section: Interpreting Individual Featuresmentioning

confidence: 99%

Supervised Learning and Model Analysis with Compositional Data

Huang¹,

Ailer²,

Kilbertus³

et al. 2022

Preprint

View full text Add to dashboard Cite

The compositionality and sparsity of high-throughput sequencing data poses a challenge for regression and classification. However, in microbiome research in particular, conditional modeling is an essential tool to investigate relationships between phenotypes and the microbiome. Existing techniques are often inadequate: they either rely on extensions of the linear log-contrast model (which adjusts for compositionality, but is often unable to capture useful signals), or they are based on black-box machine learning methods (which may capture useful signals, but ignore compositionality in downstream analyses).We propose KernelBiome, a kernel-based nonparametric regression and classification framework for compositional data. It is tailored to sparse compositional data and is able to incorporate prior knowledge, such as phylogenetic structure. KernelBiome captures complex signals, including in the zero-structure, while automatically adapting model complexity. We demonstrate on par or improved predictive performance compared with state-of-the-art machine learning methods. Additionally, our framework provides two key advantages: (i) We propose two novel quantities to interpret contributions of individual components and prove that they consistently estimate average perturbation effects of the conditional mean, extending the interpretability of linear log-contrast models to nonparametric models. (ii) We show that the connection between kernels and distances aids interpretability and provides a data-driven embedding that can augment further analysis. Finally, we apply the KernelBiome framework to two public microbiome studies and illustrate the proposed model analysis. KernelBiome is available as an open-source Python package at https://github.com/shimenghuang/KernelBiome.

show abstract

“…However, if the features are correlated, it is possible that the orderings may be biased or that permutations of one feature might result in unrealistic combinations of features and hence would cause the model to extrapolate performance [53]. For example, if all students who earned perfect scores on the physics GRE also had high GPAs, permuting GPA could cause there to be cases where a perfect physics GRE score goes with a low GPA, which would be outside of the region learned by the model.…”

Section: The Random Forest Algorithmmentioning

confidence: 99%

Rubric-based holistic review represents a change from traditional graduate admissions approaches in physics

Young¹,

Verboncoeur²,

Lam³

et al. 2021

Preprint

View full text Add to dashboard Cite

Unrestricted permutation forces extrapolation: variable importance requires at least one more model, or there is no free variable importance

Cited by 120 publications

References 32 publications

Interpretable machine learning for psychological research: Opportunities and pitfalls

Interpretable machine learning for psychological research: Opportunities and pitfalls

Supervised Learning and Model Analysis with Compositional Data

Rubric-based holistic review represents a change from traditional graduate admissions approaches in physics

Contact Info

Product

Resources

About