KNeighborsTimeSeriesClassifier not alligned with documentation

[Xhelas]

• The documentation specify that KNeighborsTimeSeriesClassifier works with variable lenght time series. It seems that it is not the case (tested with the documentation's example and other examples too.

• KNeighborsTimeSeriesClassifier does not have a one argument fit_predict method as shown in the documentation.

[rtavenar]

Hi @Xhelas ,

Thanks for raising this issue. I see two problems there:

• It does not make sense to have a single-argument fit_predict method for a supervised classification algo, so this should be changed in the docs, definitely (I guess it has to be changed here, I don't think it is referenced elsewhere)
• knn classifier with DTW similarity measure should work with variable-length time series, so here the doc is correct and we have to work on the implem

So basically, the fix should be 2-steps:

• fix the docs: should be pretty easy, but I have to figure out why latest versions of the docs do not compile on read the docs
• fix the class: I'll have a look at it asap (hopefully this afternoon)

[rtavenar]

OK, so I remove the false claims from the docs, I paste it here to put it back when the problem is solved:

Supervised classification
-------------------------

* :ref:`KNeighborsTimeSeriesClassifier <knn-clf>`

Example
~~~~~~~

.. code-block:: python

    from tslearn.neighbors import KNeighborsTimeSeriesClassifier
    clf = KNeighborsTimeSeriesClassifier(metric="dtw")
    predicted_labels = clf.fit(X, y).predict(X)

And we should definitely find a way to have kNN deal with variable length time series when (soft-)DTW is used as a metric, and similar thing for SVM+GAK. We are working on it, but it may take some time to be fixed, sorry for that.

So the remaining part of this Issue is completely related to #94 , so I suggest we move the discussion there and close this issue.