Multiple calls of .fit() has inconsistent behavior with scikit-learn

[hombit]

Currently, AADForest.fit() would do nothing when called second time, while with scikit-learn it would would cause model retraining. The same applies for fit_known() which just ignores data argument, even if it differs from data the model was previously trained with.

Here I propose to modify the Coniferest interface to add an additional method, .tune_known(known_data, known_labels). In this case:

• .fit(data) would refit every time it is called dropping all the previous training
• .fit_known(data, known_data, known_labels) would also refit
• .tune_known(known_data, known_labels) would use the same "base" (isolation forest) model and tune it for labeled data. It would fail if called before .fit or .fit_known

[matwey]

Are this names (fit_known, tune_known) exist in sklearn?

[matwey]

Related to #113

[hombit]

Are this names (fit_known, tune_known) exist in sklearn?

No, but it would be weird if .fit and fit_known behavior would be different in this way

[matwey]

Then I would propose the following alternative since it seems that having three functions is redundant:

• .fit(data, known_labels = None, known_data = None) would refit every time it is called dropping all the previous training.
• .fit_known(known_labels, known_data = None) doesn't accept data and doesn't do refit.

Mind known_labels and known_data order. If known_data is missed then known_labels are associated with data itself.

[hombit]

I would try to be duck-consistent with scikit-learn, including .fit(X, y)

[matwey]

And .fit(data, known_labels) is mostly the same as .fit(X, y).