Reproduce “Explainable Boosting Machines with Sparsity -- Maintaining Explainability in High-Dimensional Settings” on OpenResearch

About this paper

Brandon M. Greenwell, Annika Dahlmann, Saurabh Dhoble

Researchers from 84.510 and the University of Cincinnati developed a LASSO post-processing method for Explainable Boosting Machines (EBMs) to introduce sparsity, reducing model complexity by up to 95% in high-dimensional datasets while maintaining comparable predictive performance. This approach addresses the challenge of EBM interpretability and operational efficiency when dealing with a large number of features.

Key insights

• LASSO regularization, a technique known for variable selection, can be effectively adapted to prune individual feature and interaction terms within complex additive models like EBMs.
• Significant sparsity can be introduced into EBMs, drastically reducing the number of active terms, which directly enhances transparency and speeds up prediction without substantial loss in predictive accuracy.
• The approach leverages the additive structure of EBMs, allowing each component function to be individually reweighted or removed, thereby maintaining the glass-box nature of the model even after simplification.

Summary by alphaXiv.