Dataset Information

Knockoff boosted tree for model-free variable selection.

ABSTRACT:

Motivation

The recently proposed knockoff filter is a general framework for controlling the false discovery rate (FDR) when performing variable selection. This powerful new approach generates a 'knockoff' of each variable tested for exact FDR control. Imitation variables that mimic the correlation structure found within the original variables serve as negative controls for statistical inference. Current applications of knockoff methods use linear regression models and conduct variable selection only for variables existing in model functions. Here, we extend the use of knockoffs for machine learning with boosted trees, which are successful and widely used in problems where no prior knowledge of model function is required. However, currently available importance scores in tree models are insufficient for variable selection with FDR control.

Results

We propose a novel strategy for conducting variable selection without prior model topology knowledge using the knockoff method with boosted tree models. We extend the current knockoff method to model-free variable selection through the use of tree-based models. Additionally, we propose and evaluate two new sampling methods for generating knockoffs, namely the sparse covariance and principal component knockoff methods. We test and compare these methods with the original knockoff method regarding their ability to control type I errors and power. In simulation tests, we compare the properties and performance of importance test statistics of tree models. The results include different combinations of knockoffs and importance test statistics. We consider scenarios that include main-effect, interaction, exponential and second-order models while assuming the true model structures are unknown. We apply our algorithm for tumor purity estimation and tumor classification using Cancer Genome Atlas (TCGA) gene expression data. Our results show improved discrimination between difficult-to-discriminate cancer types.

Availability and implementation

The proposed algorithm is included in the KOBT package, which is available at https://cran.r-project.org/web/packages/KOBT/index.html.

Supplementary information

Supplementary data are available at Bioinformatics online.

SUBMITTER: Jiang T

PROVIDER: S-EPMC8128453 | biostudies-literature |

REPOSITORIES: biostudies-literature

ACCESS DATA

Similar Datasets

Project description:BackgroundSocial-environmental data obtained from the US Census is an important resource for understanding health disparities, but rarely is the full dataset utilized for analysis. A barrier to incorporating the full data is a lack of solid recommendations for variable selection, with researchers often hand-selecting a few variables. Thus, we evaluated the ability of empirical machine learning approaches to identify social-environmental factors having a true association with a health outcome.MethodsWe compared several popular machine learning methods, including penalized regressions (e.g. lasso, elastic net), and tree ensemble methods. Via simulation, we assessed the methods' ability to identify census variables truly associated with binary and continuous outcomes while minimizing false positive results (10 true associations, 1000 total variables). We applied the most promising method to the full census data (p?=?14,663 variables) linked to prostate cancer registry data (n?=?76,186 cases) to identify social-environmental factors associated with advanced prostate cancer.ResultsIn simulations, we found that elastic net identified many true-positive variables, while lasso provided good control of false positives. Using a combined measure of accuracy, hierarchical clustering based on Spearman's correlation with sparse group lasso regression performed the best overall. Bayesian Adaptive Regression Trees outperformed other tree ensemble methods, but not the sparse group lasso. In the full dataset, the sparse group lasso successfully identified a subset of variables, three of which replicated earlier findings.ConclusionsThis analysis demonstrated the potential of empirical machine learning approaches to identify a small subset of census variables having a true association with the outcome, and that replicate across empiric methods. Sparse clustered regression models performed best, as they identified many true positive variables while controlling false positive discoveries.

Project description:PurposeCurrent limitations in methodologies used throughout machine-learning to investigate feature importance in boosted tree modelling prevent the effective scaling to datasets with a large number of features, particularly when one is investigating both the magnitude and directionality of various features on the classification into a positive or negative class. This manuscript presents a novel methodology, "Hollow-tree Super" (HOTS), designed to resolve and visualize feature importance in boosted tree models involving a large number of features. Further, this methodology allows for accurate investigation of the directionality and magnitude various features have on classification and incorporates cross-validation to improve the accuracy and validity of the determined features of importance.MethodsUsing the Iris dataset, we first highlight the characteristics of HOTS by comparing it to other commonly used techniques for feature importance, including Gini Importance, Partial Dependence Plots, and Permutation Importance, and explain how HOTS resolves the weaknesses present in these three strategies for investigating feature importance. We then demonstrate how HOTS can be utilized in high dimensional spaces such as neuroscientific setting, by taking 60 Schizophrenic subjects from the publicly available SchizConnect database and applying the method to determine which regions of the brain were most important for the positive and negative classification of schizophrenia as determined by the positive and negative syndrome scale (PANSS).ResultsHOTS effectively replicated and supported the findings of feature importance for classification of the Iris dataset when compared to Gini importance, Partial Dependence Plots and Permutation importance, determining 'petal length' as the most important feature for positive and negative classification. When applied to the Schizconnect dataset, HOTS was able to resolve from 379 independent features, the top 10 most important features for classification, as well as their directionality for classification and magnitude compared to other features. Cross-validation supported that these same 10 features were consistently used in the decision-making process across multiple trees, and these features were localised primarily to the occipital and parietal cortices, commonly disturbed brain regions in those afflicted with Schizophrenia.ConclusionHOTS effectively overcomes previous challenges of identifying feature importance at scale, and can be utilized across a swathe of disciplines. As computational power and data quantity continues to expand, it is imperative that a methodology is developed that is able to handle the demands of working with large datasets that contain a large number of features. This approach represents a unique way to investigate both the directionality and magnitude of feature importance when working at scale within a boosted tree model that can be easily visualized within commonly used software.