Multi-Class Logistic Regression in SciKit Learn

Question

I am having trouble with the proper call of Scikit's Logistic Regression for the multi-class case. I am using the lbgfs solver, and I do have the multi_class parameter set to multinomial.

It is unclear to me how to pass the true class labels in fitting the model. I had assumed that it was similar/same as for the random forest classifier multi-class, where you pass [n_samples, m_classes] dataframe. However, in doing this, I get an error that the data is of a bad shape. ValueError: bad input shape (20, 5) -- in this tiny example, there were 5 classes, 20 samples.

On inspection, the documentation for the fit method says that the truth values are passed as [n_samples, ] -- which matches the error i'm getting -- however, I have no idea then how to train the model with multiple classes. So, this is my question: how do i pass the full set of class labels to the fit function?

i've been unable to find sample code on the Internet to model, nor this question on StackOverflow.. but i feel certain someone must know how to do it!

in the code below, train_features = [n_samples, nn_features], truth_train = [n_samples, m_classes]

clf = LogisticRegressionCV(class_weight='balanced', multi_class='multinomial', solver='lbfgs')
clf.fit(train_features, truth_train)
pred = clf.predict(test_features)

Answer 1

You seem to be confusing terms multiclass and multilabel http://scikit-learn.org/stable/modules/multiclass.html , in short:

• Multiclass classification means a classification task with more than two classes; e.g., classify a set of images of fruits which may be oranges, apples, or pears. Multiclass classification makes the assumption that each sample is assigned to one and only one label: a fruit can be either an apple or a pear but not both at the same time.

Thus data is [n_samples, n_features] and labels are [n_samples]

• Multilabel classification assigns to each sample a set of target labels. This can be thought as predicting properties of a data-point that are not mutually exclusive, such as topics that are relevant for a document. A text might be about any of religion, politics, finance or education at the same time or none of these.

Thus data is [n_samples, n_features] and labels are [n_samples, n_labels]

And you seem to be looking for multilabel (as for multiclass labels should be 1-dim). Currently, in sklearn, the only methods supporting multilabel are: Decision Trees, Random Forests, Nearest Neighbors, Ridge Regression.

If you want to learn multlabel problem with diffent model, simply use OneVsRestClassifier as a multilabel wrapper around your LogisticRegression

http://scikit-learn.org/stable/modules/generated/sklearn.multiclass.OneVsRestClassifier.html#sklearn.multiclass.OneVsRestClassifier