Sktime: Conflict between the default and documentation value of TimeSeriesForestClassifier.

[ninfueng]

Describe the bug
At the documentation of TimeSeriesForestClassifier, n_estimators = 100.
https://github.com/alan-turing-institute/sktime/blob/4dabe1afee0ad180f31c5f585f904f515ab4d0b1/sktime/classifiers/compose/ensemble.py#L48

However, at the constructor, the default value is n_estimators = 500.
https://github.com/alan-turing-institute/sktime/blob/4dabe1afee0ad180f31c5f585f904f515ab4d0b1/sktime/classifiers/compose/ensemble.py#L185

Expected behavior
All of n_estimators should be either all 100 or 500.

[mloning]

Hi, thanks for raising the issue, would appreciate a PR!

[ninfueng]

@mloning Which one should I change to 100 or 500?

[mloning]

I'd change it to 200.

[ninfueng]

Just for note, some others values also have the conflicted values as well.

max_features = None (Default), "auto" (Document)
min_impurity_split = None (Default), 1e-7 (Document)
bootstrap = False (Default), True (Document)

Please tell me, the preferred values then, I will try to pull request again.

[mloning]

These hyper-parameters go into scikit-learn's decision tree classifier, so I'd use the defaults that they have.

[ninfueng]

From sklearn documentation, DecisionTreeClassifier. Their default values are as below:

max_features = None
min_impurity_split = None

However, bootstrap is not appeared in the documentation.
Maybe we should use sktime default value, bootstrap = False, since both previous sktime default values (max_features and min_impurity_split) are the same as sklearn default values. What do you think about this?

[mloning]

The bootstrap argument comes from here: https://scikit-learn.org/stable/modules/generated/sklearn.ensemble.RandomForestClassifier.html, but happy to keep bootstrap=False

[ninfueng]

I added one more committing into Pull Request #233. I think I will end this issue now if you say so.

[mloning]

Thanks for the PR! Let me know if you're interested in contributing towards sktime more generally, we have a few interesting open issues to work on :-)