get_params and set_params for MLClassifierBase erroneous about scikit-multilearn

[niedakh]

fully merged as tests in 477efdd, thank you again @ChristianSch !

[niedakh]

Hi,

i think you are correct, what we would also need is to come up with a way to pass the parameters up/down across the classifier tree, for example lets say we have an ensemble of random ordered classifier chains. EnsembleML(RandomizeLabels(CC(DecisionTreeClassifier))).

SVC takes some params, Randomize Labels might take a seed, EnsembleML will surely take the no of chains and the size of input space subset.

Do we pass the arguments in a namespaces approach flat array, ex. the get_params from SVC would yield:

tree_parms = {'class_weight': None,
 'criterion': 'gini',
 'max_depth': None,
 'max_features': None,
 'max_leaf_nodes': None,
 'min_samples_leaf': 1,
 'min_samples_split': 2,
 'min_weight_fraction_leaf': 0.0,
 'presort': False,
 'random_state': None,
 'splitter': 'best'}

What does CC return? Do we stack this, like:

{ "cc__baseclassifierparams" : self.classifier.get_params(),
"cc__classifier": DecisionTreeClassifier(),
"cc_require_dense": True}

or do we make it a flat list

{ "cc__classifier": DecisionTreeClassifier(),
"cc_require_dense": True,
"decisiontre...classifier_criterion": ...,
...
"decisiontre...classifier_splitter": ...
}

Which one would you prefer?

I want to cleanup the MLkNN code and convert it to sparse matrices internally and that will finish the 0.0.2 feature list, then I'll look into your code ok?

[ChristianSch]

I'd look into how scikit-learn does it and then comply to this. Not sure if they have this nested params, but I'll look into it.

[ChristianSch]

Ah, I know. It's flat because the way we iterate the attributes and append attr__foobar. So flat is the way to go imho.