Timeline for Negative R2 on Simple Linear Regression (with intercept)
Current License: CC BY-SA 4.0
10 events
| when toggle format | what | by | license | comment | |
|---|---|---|---|---|---|
| Jun 2, 2023 at 11:19 | comment | added | Sextus Empiricus |
@JeanLescut The scipy.linalg.lstsq() has the same problems, but is just performing better and will make the 'error' less easily. I could make it fail with an online compiler onecompiler.com/python/3zadf7u39 (I believe that the link will take you to my scribble, although I am not sure whether it can change)
|
|
| Jun 2, 2023 at 11:12 | comment | added | Jean Lescut | Absolutely agree with you. As I wrote below, my next step would be to ask for said warning to be implemented. Despite all my trials, I couldnt make lstsq() fail on my side. I'll keep investigating. Thank you so much for your help (for me and the community !) | |
| Jun 2, 2023 at 10:37 | comment | added | Sextus Empiricus |
@JeanLescut I also had slightly different behaviour with lstsq() and changed the random seed. In your case does it also 'work' when you use np.random.seed(15)? And even when it works, the cause of the problem (not giving an error for problematic user input) has not been resolved. Ideally the function should give a warning when results are computationally unstable.
|
|
| Jun 2, 2023 at 9:01 | comment | added | Jean Lescut | You are absolutely right about CPU thing. According to my experiments, scipy.lstsq() doesnt lead to a negative R2 though. But the LinearRegression() code of sklearn convert X into a F-contiguous array before multiplying it with the parameters. This induces several small "errors" that actually becomes crazily bad in my case. I will investigate deeper and maybe propose a change in the sklearn code to keep arrays C-contiguous all along | |
| Jun 2, 2023 at 8:37 | history | edited | Sextus Empiricus | CC BY-SA 4.0 |
added 1106 characters in body
|
| Jun 2, 2023 at 8:32 | history | edited | Sextus Empiricus | CC BY-SA 4.0 |
added 1106 characters in body
|
| Jun 2, 2023 at 7:25 | comment | added | Sextus Empiricus |
@JeanLescut I know that lm in R would provide an error telling that the matrix is singular. I recently encountered a similar question for MATLAB that had different behaviours with a similar issue of matrix inversion. stats.stackexchange.com/questions/616591
|
|
| Jun 2, 2023 at 6:56 | comment | added | Jean Lescut | Amazing, we're getting closer ! My hope here is to try to implement a check in lstsq() and issue a warning when needed. It's understandable that the function fails, but not that it fails silently... | |
| Jun 1, 2023 at 23:59 | history | edited | Sextus Empiricus | CC BY-SA 4.0 |
added 37 characters in body
|
| Jun 1, 2023 at 23:52 | history | answered | Sextus Empiricus | CC BY-SA 4.0 |