Only one letter different from GBM, GLMs (generalized linear models) take a very different approach. Whereas decision trees are based on logic, and deep learning is a black box inspired by the human brain, GLMs are based on mathematics. The underlying idea is something you almost certainly did at school: make a scatterplot of data points on graph paper, then draw the best straight line through them. And perhaps you have used
lm() in R or
linear_model.LinearRegression in Python’s scikit-learn, or something similar, to have the computer do this for you. Once you progress beyond the graph paper you can apply it to any number of dimensions: each input column in training data counts as one dimension.
Sticking with school memories, when I first heard about Einstein’s general and special theories of relativity, I assumed the special theory was the complicated one, to handle some especially difficult things that the general-purpose one couldn’t deal with. It turns out the general theory was called that because it generalized both the special theory and some other stuff into one über-complicated theory. And so it is with generalized linear models: they can do your grandfather’s linear model (in fact, that is the default behavior), but they can also do other stuff.
That other stuff comes down to a couple of things: using
link(y) = mx + c instead of
y = mx + c (where
link() is a function that allows introducing nonlinearity); and specifying the distribution of the ...