Preface
Additive modeling algorithms are an essential part of the machine learning modeler’s tool kit when working with tabular data. Such predictive models have excellent intrinsic interpretability and often have optimal or near optimal performance. While linear additive models are the most common type of additive models, and they often have optimal performance, there are also many cases where additive functions need more flexibility.
To address this need we have developed the Model Sculpting method for building additive models. This is a model building pipeline that starts from developing a strong learner, such as boosted trees, and then extracts an additive model that best approximates the strong learner, thereby providing a more interpretable model with limited cost in performance.
The slide deck below provides an overview of the Model Sculpting method and its applications. The modsculpt R package provides an implementation of the method and the example workflow demonstrates how to use the package to build an additive model.
Contents
Citation
@misc{friesenhahn2024,
author = {Michel Friesenhahn and Ondrej Slama and Kenta Yoshida},
title = {The {Unreasonable} {Effectiveness} of {Additive} {Models}},
date = {2024-09-20},
url = {https://go.roche.com/stats4datascience},
langid = {en}
}