ConfusionMatrix classes: one label must be specified in y true.

When trying to passes class names to the ConfusionMatrix, I’m getting the following error:

/usr/local/lib/python3.6/site-packages/sklearn/metrics/classification.py:248: FutureWarning: elementwise comparison failed; returning scalar instead, but in the future will perform elementwise comparison
  if np.all([l not in y_true for l in labels]):
Traceback (most recent call last):
  File "rc.py", line 63, in <module>
    viz.score(X_test, y_test)
  File "/Users/benjamin/Repos/ddl/yellowbrick/yellowbrick/classifier/confusion_matrix.py", line 107, in score
    y, y_pred, labels=self.classes_, sample_weight=sample_weight
  File "/usr/local/lib/python3.6/site-packages/sklearn/metrics/classification.py", line 249, in confusion_matrix
    raise ValueError("At least one label specified must be in y_true")
ValueError: At least one label specified must be in y_true

This is do to the fact that the y_true vector is label encoded (e.g. ones and zeros) but that the labels are strings, e.g. “occupied” and “unoccupied”.

Here is the code to produce the error:

df = pd.read_csv('examples/data/occupancy/occupancy.csv')

features = ["temperature", "relative humidity", "light", "C02", "humidity"]
target = "occupancy"

X = df[features]
y = df[target]
classes = ["unoccupied", "occupied"]

X_train, X_test, y_train, y_test = tts(X, y, test_size=0.2)

viz = ConfusionMatrix(MultinomialNB(), classes=classes)
viz.fit(X_train, y_train)
viz.score(X_test, y_test)
viz.poof()

Note that:

viz = ConfusionMatrix(MultinomialNB())

Does not raise an exception.

[cjmorale]

@bbengfort Is anyone working on this bug yet? I was able to reproduce the error last night on my local environment and wanted to offer my help.

[bbengfort]

@rebeccabilbro yep, sorry about this, it should have been closed on June 8 with #253. For future reference, the fix to this is as follows:

viz = ConfusionMatrix(
    MultinomialNB(), classes=classes, label_encoder={0: "unoccupied", 1: "occupied"}
)

Where the label_encoder argument can be a sklearn.preprocessing.LabelEncoder (or really anything with inverse_transform that does the mapping) or a dict.

closed via ff10ca0