scikit-learn: .predict_proba() for SVC produces incorrect results for binary classification

Description

svm.predict_proba() produces revered results for binary classification

• this seems to be specific to binary classification. For example, it works fine for 3 way classification, which is in the test: https://github.com/scikit-learn/scikit-learn/blob/master/sklearn/svm/tests/test_svm.py#L307
• I think this is related to #394, but in this case we are using the same svm object so I think predict and predict_proba should agree
• Acknowledgement: @amennen noticed this error 👍

Steps/Code to Reproduce

here’s the code on colab: https://colab.research.google.com/github/qihongl/random/blob/master/sklearn-svm-predict-proba-bug.ipynb

# train a svm 
X = np.array([[-1, -1], [1, 1]])
y = np.array([0, 1])
svm = SVC(probability=True)
svm.fit(X, y) 

# SVM makes reasonable prediction on the learned examples... 
print(svm.predict(X))

# but it makes the reversed probability estimates... 
print(svm.predict_proba(X))

Expected Results

[0 1]
[[0.66383953 0.33616047]
 [0.33916469 0.66083531]]

# i.e. when the prediction is class 0, prob(class0) should be bigger than prob(class1)

Actual Results

[0 1]
[[0.33616047 0.66383953]
 [0.66083531 0.33916469]]

# i.e. when the prediction is class 0, prob(class0) < prob(class1)

Versions

0.20.3