There is no simple rule how to properly generate baselines for all methods.

Especially in XPS, the baseline selection may be quite controversial. There the Shirley baseline is nowadays considered sort of a "minimum standard", but of course it may not account for enough. The CasaXPS manual is very helpful on this, even if you use a different software:

http://www.casaxps.com/help_manual/manual_updates/peak_fitting_in_xps.pdf

In Raman, I've seen most people using rather primitive linear background approaches, at least as long as the spectrum is not a big forest.

There are no rules. For XRD measurements you can use polynomial functions with different degree or also manually provide points that the software can interpolate...the choice is not simple

Speaking in general terms, a baseline subtraction is removing whatever contribution your data has that is completely irrelevant to your data analysis. I can only say something about Raman, but the best way to subtract a baseline is to measure your sample and your substrate using the same laser power, number of acquistions and time and subtract, from the sample signal, the signal from your substrate. If your sample has some contribution coming from some sort of luminescence, one way to remove the luminescent background would be to measure your sample in a parallel polarization configuration and follow that with a cross polarization configuration. Generally, the PL is unpolarized, and your measurement in cross polarization should remove the totally symmetric Raman modes, leaving only the other types of modes and the PL, which you can use to remove the PL from the parallel configuration. Again, if you know enough about Raman, and even PL, you'll know that these are suggestions rather than rules, because there is not a one perfect rule for all cases. As with most people though, you'll eventually have enough experience to do background subtraction easily.

Hope that helps.