Why do ELI5 probabilities differ from predict_proba

Question

I have an sklearn pipeline consisting of a TfidfVectorizer and a SGDClassifier(loss=\'log\'), fitted on a multi-label training dataset. When I then