I have a dataset
data that has missing values. I am trying two ways of imputing these values, but I would like to compare them.
In the first method I am using a combination of techniques. Some are filled with the mean, others with conditional mapping, and a couple with models fit to other features. I then call this
In the second method I am putting in the original
data and using sklearn's
IterativeImputer in a pipeline.
Both methods will be used to fit a RandomForestClassifier of the same structure. If I use the same
random_state in the splits of
data, will this give me a good comparison of imputation techniques? e.g. will
train_test_split(imp_data[inputs], imp_data[target], random_state=1)
train_test_split(data[inputs], data[target], random_state=1)
produce directly comparable results?