This is a section of code copied from some software I use:
# Group Splitting for train & Val (Dev)
if not is_predict:
gss_train_val = GroupShuffleSplit(n_splits=1,
test_size=(1/4),
random_state=42).split(train_dataset,
groups=train_dataset['id'])
X_train_inds, X_val_inds = next(gss_train_val)
training_data = train_dataset.iloc[X_train_inds]
val_data = train_dataset.iloc[X_val_inds]
training_data.reset_index(drop=True)
val_data.reset_index(drop=True)
# Group Splitting for train & test (Dev)
gss_train_test = GroupShuffleSplit(n_splits=1,
test_size=(1/Vars.FOLDS),
random_state=42).split(training_data,
groups=training_data['id'])
X_train_inds, X_test_inds = next(gss_train_test)
train_data = training_data.iloc[X_train_inds]
test_data = training_data.iloc[X_test_inds]
train_data.reset_index(drop=True)
test_data.reset_index(drop=True)
What I am curious about is varying the test data size to see the effect on accuracy, and also altering the random_state value. E.G. if I alter test_size=(1/4) to test_size=(1/3), does this effectively test a bigger data sample, and potentially increase accuracy? Am I likely to break anything by messing about with these numbers; as long as I note the original script, I should be able to re-run the software and be back where I started? Thanks in advance for any help