I am trying to understand what type of shift(s) exist in my problem to get a better grasp. I have a dataset which comprises of a deep neural network's (DNN) runtime latency ($y$), its architecture ($a$) as well as the hardware it was run on ($h$). The runtime of a DNN depends on $a$ and $h$ and can be formulated as $f(a,h;\theta) \rightarrow y$.
Now let's say I 100 different DNN models executed on $h_1$ and the same 100 models also executed on $h_2$. Domain adaptation typically relies on the assumption of covariate shift i.e. the input features (distribution) changes. The classic example is SVHN and MNIST datasets.
But in this case, the DNNs - which are encoding in some form to become an input feature - are the same between both distributions. $h_1$ and $h_2$ certainly change but only 2 values exist for them.
The output distribution changes significantly, however! For example, models that execute on $h_1$ exhibit a far lower DNN runtime than when run $h_2$. What type of shift in the domain is this?
As I deep deeper into the literature, a lot of paper make the assumption of covariate shift. But that assumption is probably not true for my case. Thats why I want to know, what does the literature term a shift in the
