Per introduced in the lecture

“Prior probability shift is basically the opposite of covariate shift. The distribution of your labels changes, but your input data stays the same. Concept drift can be thought of as a type of prior probability shift”

While per the lecture in Course 2: *Detecting Data Issues*

the definition of concept shift should be Ptrain(y|x) != Pserv(y|x) but Ptrain(x) = Pserv(x), seems it is exact the same with the prior probability shift defined here.

BTW, just want to confirm if the term shift and drift interchangeable? Since saw many concept shift, concept drift, dataset shift, data drift … etc present in this specialization.

Any progress on this post since 6 days no response?

As far your comment on covariate and concept shifts are concerned, please help me understand what `here`

refers to.

Shift & drift:

I’ve seen external sources use shift and drift interchangeably. I’m assuming that they consider change across 2 datasets (be it same dataset across time or different datasets) as a shift in distribution.

Drift refers to changes in the dataset involving a single split (e.g. training data) collected over time.

Skew refers to difference in properties of 2 dataset splits (e.g. training and serving) at a point in time. The term `* shift`

(Covariate, Dataset, Concept), as per the lectures is closely aligned with `skew`

.

Is that a easier to understand it like shift is a super set of drift and the drift concept focusing on the change per split overtime on the same dataset (means beside the sample timestamp all the rest are identical), while shift also involving changes caused by other factors like different source, data preprocessing beside the drift covered?

I meant the formula introduced in course 2

`Ptrain(y|x) != Pserv(y|x) but Ptrain(x) = Pserv(x)`

and the mentor’s describe about **Prior probability shift** here are the same issue, but the mentor also said

“Concept drift can be thought of as a type of prior probability shift”

I think it hint there are other kinds of prior probability shift other than concept shift, want to know what that is.

To on my previous reply, using the word `shift`

seems quite generic to me.

Since you now know what each term means, please consider the following:

- Use
`drift`

when talking about the same dataset split.
- Use
`* shift`

when talking about 2 different splits of a dataset.
- If you’re unsure, explain your scenario in terms of distributions across 2 splits and hopefully the existing types of shift cover your case. If not, you might just stumble across a new way of comparing splits.

`Deep learning specialization`

goes over methods to deal with differences across splits. Refreshing your memory on courses 2 and 3 will help.

I find the shifts mentioned in the slides to be the complete in terms of coverage. Do use external sources to explore further.

I see, thus I can treat Prior Probability Shift as an alias of concept shift right?