- What is the difference(s) between boosted regression trees (BRT) and generalized boosted models (GBM)? Can they be used interchangeably? Is one a specific form of the other?
- Why did Ridgeway use the phrase "Generalized Boosted Regression Models" (GBM), to describe what Friedman had previously proposed as "Gradient Boosting Machine" (GBM)? These two acronyms are identical, describe the same thing, but are derived from different phrases.
I am having trouble determining how the terms BRT and GBM differ. From what I understand both are terms to describe classification and regression trees that have stochasticity incorporated through some sort of boosting (e.g. bagging, bootstrapping, cross-validation).
Also, from what I gather the term GBM was first coined by Friedman (2001) in his paper "Greedy function approximation: a gradient boosting machine". Ridgeway then implemented the procedure described by Friedman in 2006 in his package "Generalized Boosted Regression Models" (GBM). In my field (ecology) Elith et al. (2008) was the first to demonstrate Ridgeway's
gbm package for species distribution modelling. However, the authors in Elith et al. use the term "boosted regression tree" (BRT) to describe Friedman and Ridgeway's GBM theory and implementation.
I am confused as to if these terms can be used interchangeably? It is somewhat confusing that one author would use the same acronym (from a different phrase) to describe the same theory that a previous author proposed. It is also confusing that the third author used a completely different term when describing this theory in ecological terms.
The best I can come up with is that BRT are a specific form of GBM in which the distribution is binomial, but I am not sure of this.
Elith et al. define boosted regression trees like this… "Boosted regression trees combine the strengths of two algorithms: regression trees (models that relate a response to their predictors by recursive binary splits) and boosting (an adaptive method for combining many simple models to give improved predictive performance). The final BRT model can be understood as an additive regression model in which individual terms are simple trees, fitted in a forward, stagewise fashion" (Elith et al. 2008).