Interpretability in Generalized Additive Models

S. N. Wood, Y. Goude, M. Fasiolo*

*Corresponding author for this work

Research output: Chapter in Book/Report/Conference proceedingChapter

Abstract / Description of output

Modelling the effect of a covariate vector on the distribution of a response variable, requires some structural assumptions, if the curse of dimensionality is to be avoided. Generalized additive models (GAMs) assume that the effects of the covariates are additive, with no, or only low order, interactions between effects. The additive assumption ensures scalability in the number of covariates and facilitates computational efficiency during model fitting. It also enhances model interpretability, which is critically important during model building and checking, as well as for communicating modelling results. This chapter formally introduces standard GAMs, as well as more flexibile GAMs for location shape and scale (GAMLSS). Is also shows how to interactively build and improve GAM and GAMLSS models via the mgcv and mgcViz R packages, which exploit their modular and interpretable structure. The final part of the chapter shows how to exploit the additive structure of GAMs to build powerful predictive models, by using random forests and online aggregation methods.

Original languageEnglish
Title of host publicationInterpretability for Industry 4.0: Statistical and Machine Learning Approaches
EditorsAntonio Lepore, Biagio Palumbo, Jean-Michael Poggi
PublisherSpringer
Pages85-123
Number of pages39
ISBN (Electronic)9783031124020
ISBN (Print)9783031124013
DOIs
Publication statusPublished - 1 Jan 2022

Fingerprint

Dive into the research topics of 'Interpretability in Generalized Additive Models'. Together they form a unique fingerprint.

Cite this