Correct me if I am wrong, oob_score is the score of random forest regressor on that validation set created by the unions of all the rows left-out after each tree training set.
Yes, I guess. When you take a sample with replacement, some rows will be left out during the training phase of each tree. Therefore we can think of getting those rows and compute the score with those. I do not know if it is appropriate to define it a ‘validation set’, though!