I’m in a place where we’re unable to train models on data due to GDPR. What I want is to predict people getting a job (y) given (x,x,x,x…) their employment type working full time or part time, work sector, area and age.
Since I cannot train on real data, I created an aggregate from the real data. Then used numpy.random.choice to generate synthetic data frame with data from the aggregate. Then created a decision tree boost model that had a good f1 score and accuracy. The behaviour of the model is similar to the aggregate.
When I presented this to people, they informed me that what I did is not a real model and by recreating data from an aggregate, you loose important correlations from the data for individuals represented in the dataset.
I disagree on these points since the model is acting based on the aggregate.
After reading, I figured out that you do not need to have correlations to build models. There are many research papers based on non linear dataset with no/few correlations.
I don’t understand how a model from an aggregate looses its statistical traits.