Making ROC curve using python for multiclassification

Question

Following up from here: Converting a 1D array into a 2D class-based matrix in python

I want to draw ROC curves for each of my 46 classes. I have 300 test samples for which I've run my classifier to make a prediction.

y_test is the true classes, and y_pred is what my classifier predicted.

Here's my code:

    from sklearn.metrics import confusion_matrix, roc_curve, auc
    from sklearn.preprocessing import label_binarize
    import numpy as np

    y_test_bi = label_binarize(y_test, classes=[0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18, 19,20,21,2,23,24,25,26,27,28,29,30,31,32,33,34,35,36,37,38,3,40,41,42,43,44,45])
    y_pred_bi = label_binarize(y_pred, classes=[0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18, 19,20,21,2,23,24,25,26,27,28,29,30,31,32,33,34,35,36,37,38,3,40,41,42,43,44,45])
    # Compute ROC curve and ROC area for each class
    fpr = dict()
    tpr = dict()
    roc_auc = dict()
    for i in range(2):
        fpr[i], tpr[i], _ = roc_curve(y_test_bi, y_pred_bi)
        roc_auc[i] = auc(fpr[i], tpr[i])

However, now I'm getting the following error:

Traceback (most recent call last):
  File "C:\Users\app\Documents\Python Scripts\gbc_classifier_test.py", line 152, in <module>
    fpr[i], tpr[i], _ = roc_curve(y_test_bi, y_pred_bi)
  File "C:\Users\app\Anaconda\lib\site-packages\sklearn\metrics\metrics.py", line 672, in roc_curve
    fps, tps, thresholds = _binary_clf_curve(y_true, y_score, pos_label)
  File "C:\Users\app\Anaconda\lib\site-packages\sklearn\metrics\metrics.py", line 505, in _binary_clf_curve
    y_true = column_or_1d(y_true)
  File "C:\Users\app\Anaconda\lib\site-packages\sklearn\utils\validation.py", line 265, in column_or_1d
    raise ValueError("bad input shape {0}".format(shape))
ValueError: bad input shape (300L, 46L)

Answer 1

roc_curve takes parameter with shape [n_samples] (link), and your inputs (either y_test_bi or y_pred_bi) are of shape (300, 46). Note the first

I think the problem is y_pred_bi is an array of probabilities, created by calling clf.predict_proba(X) (please confirm this). Since your classifier was trained on all 46 classes, it outputs a 46-dimensional vectors for each data point, and there is nothing label_binarize can do about that.

I know of two ways around this:

1. Train 46 binary classifiers by invoking label_binarize before clf.fit() and then compute ROC curve
2. Slice each column of the 300-by-46 output array and pass that as the second parameter to roc_curve. This is my preferred approach by I am assuming y_pred_bi contains probabilities