Unbiased Measurement of Feature Importance in Tree-Based Methods
arXiv:1903.05179
Abstract
We propose a modification that corrects for split-improvement variable importance measures in Random Forests and other tree-based methods. These methods have been shown to be biased towards increasing the importance of features with more potential splits. We show that by appropriately incorporating split-improvement as measured on out of sample data, this bias can be corrected yielding better summaries and screening tools.
arXiv commentadd Section 3.4 to compare with other methods for dealing with similar bias; add more simulation results in Section 5; add link to Github repository for code access
References in corpus (11)
- Scikit-learn: Machine Learning in Python
- Correlation and variable importance in random forests
- Consistency of random forests
- Improving the precision of classification trees
- Narrowing the Gap: Random Forests In Theory and In Practice
- Quantifying Uncertainty in Random Forests via Confidence Intervals and Hypothesis Tests
- Predicting the direction of stock market prices using random forest
- A Debiased MDI Feature Importance Measure for Random Forests
- Boosting Random Forests to Reduce Bias; One-Step Boosted Forest and its Variance Estimate
- Approximation Trees: Statistical Stability in Model Distillation
- Boulevard: Regularized Stochastic Gradient Boosted Trees and Their Limiting Distribution
Cited by in corpus (5)
- Unbiased variable importance for random forests
- Boosting Random Forests to Reduce Bias; One-Step Boosted Forest and its Variance Estimate
- Hybrid analytic and machine-learned baryonic property insertion into galactic dark matter haloes
- Provable Boolean Interaction Recovery from Tree Ensemble obtained via Random Forests
- ControlBurn: Feature Selection by Sparse Forests