I understand that, in the context of a binary classification problem, downsampling the majority class is a useful strategy to come up with a smaller, computationally friendly dataset. Using this modified dataset would not improve the accuracy of whatever model is trained on it: if the downsampling of the majority class in the original dataset is done carefully (accounting for its statistical characteristics), training a model on the downsampled dataset would give us a model with a slightly degraded accuracy (as compared with the accuracy obtained training the same model on the original dataset). The only thing that will improve accuracy in these circumstances is to find new features that provide a clear separation between classes.
Can someone please clarify what advantage is provided by upsampling the minority class in this (or in a different) context?
This answer states "we will apply the model either to the full dataset, or to an oversampled balanced one, which contains all the instances of the rare class and the same number of samples from the majority class (so the oversampled dataset is smaller than the full dataset)." This is completely different from the situation I describe in my question: to the best of my understanding, the term "oversampling" means artificially generating instances of the minority class, thus creating a balanced dataset with a larger number of samples than the original one (I am using the term "upsampling" to describe this same procedure).