For the same test set, two different seeds are used for two experiments. The training sets of these two experiments are both larger, but the result of one experiment is that the accuracy of the test set increases, and the result of the other experiment is that the accuracy of the test set decreases. Although the training sets of the two sets of experiments are larger, they are not an extension of the small data set, but re-randomly sampled a larger data set. However, the training sets of these two experiments are the same.
This phenomenon is very interesting. A larger training set makes it easier to understand the accuracy of the test set, but why does a larger training set also cause the accuracy of the test set to decrease? Perhaps it is because the larger training set contains more different distributions than the test set.
