Linear Regression - Statistics vs Data Mining

[ManUtd]

Hi Professor Yaser/Everyone-

A question about looking at regression from a stats vs data mining angle.

Stats - checks for correlated variables, normality of residuals/variables (non-linear transformations probably take care of this), homoscedasticity etc.

Data Mining - as you had mentioned, we want to keep it general.

Does that mean -
a) we don't care about these assumptions or we do care, but they come into play later on.
b) we are at a higher risk for getting misleading results.

It would be nice to have your thoughts on this.

Thanks,
Kartik

[htlin]

Originally Posted by ManUtd

Hi Professor Yaser/Everyone-

A question about looking at regression from a stats vs data mining angle.

Stats - checks for correlated variables, normality of residuals/variables (non-linear transformations probably take care of this), homoscedasticity etc.

Data Mining - as you had mentioned, we want to keep it general.

Does that mean -
a) we don't care about these assumptions or we do care, but they come into play later on.
b) we are at a higher risk for getting misleading results.

It would be nice to have your thoughts on this.

Thanks,
Kartik

IMHO, from an ML perspective, we care about the out-of-sample performance. Different levels of assumptions gives you different guarantees on the out-of-sample performance. Statisticians (mathematicians) are often willing to go for more assumptions to get more decent mathematical guarantees, including out-of-sample performance and other measures of mathematical interests. The problem with assumptions is that they may or may not be realistic. So understanding the guarantee when using the least assumptions is important and is an angle we are presenting in the book.