Recommendations for setting data_norm and bounds in models after StandardScaler

[PaulineMauryL]

Hi,
Thank you very much for this great work!

I have been experimenting with DiffPrivLib using pipelines such as:

dpl_pipeline_with_scaler = Pipeline([
    ('scaler', models.StandardScaler(epsilon = 0.5, bounds=dpl_bounds)),
    ('classifier', models.LogisticRegression(epsilon = 1.0, data_norm=log_reg_data_norm))
])

and

kmeans_pipeline = Pipeline([
    ('scaler', models.StandardScaler(epsilon = 0.5, bounds=dpl_bounds)),
    ('kmeans', models.KMeans(n_clusters = N_CLUSTERS, epsilon = 5.0, bounds=([-3, -3, -3, -3, -3], [3, 3, 3, 3, 3]))),
])

My question is about how best to set the parameters in the model following the standard scaler:

• data_norm (for LogisticRegression for instance)
• bounds (for KMeans for instance)

For now, I’ve been reasoning as follows:

• After applying StandardScaler, most values should empirically fall in [-3, 3] per feature (using the 68–95–99.7 rule - Wikipedia) (assuming I have a normal distribution..).
• Based on this, I set KMeans bounds to [-3, 3] per feature.
• For the LogisticRegression, I estimate log_reg_data_norm = np.sqrt(n_features * (3**2)).

However, this approach may not be ideal. Do you have any recommendations for better ways to set these parameters after scaling? Or did I misunderstood and should instead use the original input data bounds (as used in StandardScaler)?

Thanks again for your work on this library!

[PaulineMauryL]

Hi again, I found a solution using the model's internal parameters .mean_ and .var_ and computing the new theoretical bounds with it (and using a similar method for bounds after PCA).
Unfortunately, if I use a pipeline, it is still not exact as the .fit() on pipelines retrain all models so internal parameters may vary a bit.