Scikit-learn: DecisionTreeClassifier unknown label type: 'continuous-multioutput'

You should be using DecisionTreeRegressor

Again, documentation lacks information on how many classes can classification handle. I can see that my dataset has waaaay too many classes, but your error message mentioned something like 'labels' which was confusing enough to forget how the dataset actually look like and meddle with methods of passing datasets.
I've updated the issue and ask you to reopen it.

Classification targets should be represented as integers or as strings. You can ask Pandas to read the target data in as a string and you'll be fine.

Or use a DecisionTreeRegressor

That's not my problem

See 'Expected Results' section of my issue

You're right that the error message could be more useful, but the documentation for fit does say "class labels in classification". Feel free to submit a clearer issue about needing to document the expected data type for classification ys, and another for raising appropriate error messages when float data is passed as y to a classifier.

Let me cite the whole section of documentation documenting parameter y of function fit in class DecisionTreeClassifier

The target values (class labels in classification, real numbers in regression). In the regression case, use dtype=np.float64 and order='C' for maximum efficiency.

That does not say that classes have a cap. What makes a target variable labeled continuous? How many classes have to be there to be considered regression type target variable? If it sais about regression, then can I do regression using DecisionTreeClassifier? Why not? Etc...

As for your previous comment:

Classification targets should be represented as integers or as strings. You can ask Pandas to read the target data in as a string and you'll be fine.

Does that mean that classes can't be represented as floats? Or as dicts? Lists? Tuples? Longs? Doubles? bytes? I know it is logical to represent classes as integers or strings, since they should not be plenty. But do they have to? What are the limitations?

And as to creating new ticket, isn't that useless since we've had quite a talk in here? Creating new ticket just to explain other guy the same thing?

It's not number of classes. It's use of non-integers and non-strings.

I like the issue descriptions to be focused. Your concern as raised here
seemed to be more of a usage problem.

And please don't hassle me about what I suggest. This isn't the only issue
I'm dealing with.

On 2 November 2016 at 00:24, Piotr Kamoda [email protected] wrote:

Let me cite the whole section of documentation documenting parameter y of
function fit in class DecisionTreeClassifier

The target values (class labels in classification, real numbers in
regression). In the regression case, use dtype=np.float64 and order='C' for
maximum efficiency.

That does not say that classes have a cap. What makes a target variable
labeled continuous? How many classes have to be there to be considered
regression type target variable? If it sais about regression, then can I do
regression using DecisionTreeClassifier? Why not? Etc...

And as to creating new ticket, isn't that useless since we've had quite a
talk in here? Creating new ticket just to explain other guy the same thing?

—
You are receiving this because you modified the open/close state.
Reply to this email directly, view it on GitHub
https://github.com/scikit-learn/scikit-learn/issues/7801#issuecomment-257565248,
or mute the thread
https://github.com/notifications/unsubscribe-auth/AAEz63zNA0Qc4lzgHttXx-4VFkJFwFaDks5q5z16gaJpZM4KlSFw
.

You don't have to solve it today, I'm only trying to make the issue of bad error descriptions and bad documentation on tree classifier and regressor to become active and a task for future releases.

For the error message would "Unsupported output type: 'continuous-multioutput'" be better? That is the _real_ issue. Also see #7809 for the docstring.

That's better. But still I don't understand why you won't name it as it is. Because literature mostly calls that 'Target' variables, and output could be mistaken with function output. Exception was thrown from function 'check_classification_targets', so even you say that's 'target' variable, and still you want to call it 'label' or 'output'. I'm not a member of scikit-learn member, so you will do as you please, but I would recommend to use words 'Target variable' in doscstring and error message. And I ask you to describe anywhere rules that input data (or target) should follow. A short sentence - 'Target variable (parameter y) has to be int or str'.

Maybe it's worth mentioning in/alongside the new section (45cb11d / #7519)
on multiclass and multilabel fitting in the tutorial. Or maybe this all
belongs in a section of the user guide on data representation conventions,
describing input/output formats for all standard methods...?

On 2 November 2016 at 20:56, Piotr Kamoda [email protected] wrote:

That's better. But still I don't understand why you won't name it as it
is. Because literature mostly calls that 'Target' variables, and output
could be mistaken with function output. Exception was thrown from function
'check_classification_targets', so even you say that's 'target' variable,
and still you want to call it 'label' or 'output'. I'm not a member of
scikit-learn member, so you will do as you please, but I would recommend to
use words 'Target variable' in doscstring and error message. And I ask you
to describe anywhere rules that input data (or target) should follow. A
short sentence - 'Target variable (parameter y) has to be int or str'.

—
You are receiving this because you modified the open/close state.
Reply to this email directly, view it on GitHub
https://github.com/scikit-learn/scikit-learn/issues/7801#issuecomment-257820087,
or mute the thread
https://github.com/notifications/unsubscribe-auth/AAEz68AqYGWrP1C-BinLXGGHNt0VtV_qks5q6F5rgaJpZM4KlSFw
.

I've scanned the document and it seems a good place to mention those conventions. Also if you don't want to obfuscate the error messages too much then idea of putting that information in user guide isn't bad as well.
Well, the final solution (if any) will be as you wish it to be, I'm just saying that the idea seems ok, but you have your conventions. I wont make you do something.

'Target variable (parameter y) has to be int or str'. is not right, because we support multi-label and multi-output multi-target

Also, arbitrary objects that are not floats are supported as class labels, they don't have to be integers or strings.

If we put as imput training_data_X, training_scores_Y to fit method it cause error. To avoid it we will convert and encode labels

from sklearn import preprocessing
from sklearn import utils
lab_enc = preprocessing.LabelEncoder()
y_train = lab_enc.fit_transform(y_train)
print(y_train)
print(utils.multiclass.type_of_target(y_train))
print(utils.multiclass.type_of_target(y_train.astype('int')))
print(utils.multiclass.type_of_target(y_train))

I'm having this same issue, is there a fix for it?