Dear Luiz,

You can use dB or dB(A), but if you are trying to correlate subjective results with objective data, then I would recommend touse psyco-acosutic parameters such as loudness.

It depends on what the regression is trying to achieve, and if there is an applicable criteria you are trying to meet.  I expect this is for human hearing, so dBA is a typical noise metric that is used.  However, the type of noise measurement (Leq, Lmax, L90, etc.) should also be considered.

Perhaps a bit more on the background of your project would help.

Frank is right. Probably we should know more about your problem. May be you want to correlate annoyance, level perception, comfort,etc. Thus sound quality is a different problem from sound annoyance. This will dictate what parameters you should use

First of all, thank you for the help. 

However, I think I didn't specify very well the problem I am addressing.

Actually, the regression problems I am tackling have no particular domain. I am working with symbolic regression and testing a GP method with datasets from the UCI machine learning repository (and other sources). However these datasets have no direct relation with the signal processing field.

For those who are following this thread, I've recently proposed regression versions of metrics for measure noise in classification problems.

My work will appear in GECCO'17. If you are interested, it should be available in the next months. The paper title is "How Noisy Data Affects Geometric Semantic Genetic Programming".

Best regards.