Hello, I have a problem with Dataiku's predictions, I don't know how they work. I am creating a model to get sales predictions. So far no problem, but when I try to reproduce the model, using the same variables, the same algorithm, the same file structure, the result is different from the one I had initially .The R² and the mean squared error changes also, especially with Random forest and XGboost. I don't understand why, is this normal?
The most important one by far is that Dataiku preprocesses your data. This includes numerous things, but on "normal" data, the two most important ones are dummifying categorical values, and standardizing numericals.
In order to get as close as possible to the results of Visual ML using external code, you'd need to disable as much as possible of it, by:
Only passing numerical variables to the Visual ML
Disabling standardization on all numerical variables in the Visual ML
Another thing that may be different is train/test split. In order to get more reproducible results, you'd want to pass explicit train and test sets to both Visual ML and your code
Then there is hyperparameter search. The easiest here would be to disable it entirely.