I want to determine whether my assumption that the dataset I'm using is i.i.d. is in fact valid (for an arbitrary dataset, perhaps made of images). I have done quite a bit of research already, looked through various papers on google scholar, but haven't found what I'm looking for. Is there any half-decent measure of independence of a dataset? Even if I can't guarantee it, are there any common modern methods used to test for this assumption?
$\begingroup$
$\endgroup$
6
-
1$\begingroup$ Since you bring up independence and identically distributed, you're working with the idea that the samples, using your example of images, are randomly produced, correct? $\endgroup$– A rural readerCommented Nov 2, 2023 at 0:01
-
1$\begingroup$ Welcome to Cross Validated! You want to know about the independence of what? $\endgroup$– DaveCommented Nov 3, 2023 at 15:34
-
$\begingroup$ i.i.d. is usually treated as a theoretical assumption which rarely, if ever, holds with empirical data. I, too, would be interested in a test of this assumption. $\endgroup$– user78229Commented Nov 3, 2023 at 15:44
-
2$\begingroup$ Why do you want to determine whether your dataset meets the i.i.d. assumption? What will you do if it does, and what will you do if it doesn't? One way an image dataset could violate this assumption is if the images were taken by different people, with each person taking multiple similar pictures. Or if the pictures were taken in some order over time, and pictures taken nearby in time are more highly correlated. In these cases you may want to modify your train/test strategy. $\endgroup$– AdrianCommented Nov 3, 2023 at 15:47
-
$\begingroup$ If my comment above feels relevant, you may be interested in stats.stackexchange.com/questions/564063/… $\endgroup$– AdrianCommented Nov 3, 2023 at 15:49
|
Show 1 more comment