Seems like a simple problem:
We are collecting feedback from users on X. X is generated by a machine learning model. X follows a clear schema and is shown to the users after some transformations.
When we update the transformations or the schema, we plan to migrate the old feedback data to make use of it.
But there can be both semantic and visual drifts as a result of the changes.
What are the best practices here? Do you migrate training data? How do you account for the biases resulting from the migration?
