# SciKit-Learn: predict values sometimes different from top predict_proba entries?

I noticed that in my text classification problem, I get significantly worse results when I use the max of predict_proba's output as compared to the straight predict value. Digging further, I found that the highest entries in each row of predict_proba don't always have index equal to the entry of predict. Example:

> [*test.target]
[61, 13, 11, 89, 11, 71, 118, 33, 52, 57, 16, 57, 100, 24, ...]
> [*text_clf.predict(test.data)]
[61, 16, 11, 16, 11, 89, 26, 33, 16, 57, 16, 118, 11, 26, ...]
> [max(enumerate(prob), key=lambda p: p[1])[0] for prob in text_clf.predict_proba(test.data)]
[60, 16, 11, 16, 11, 87, 25, 32, 16, 56, 16, 115, 11, 25, ...]


You can see that predictions roughly follows test.target (not the greatest accuracy, but cest la vie), and predict_proba roughly follows predict, however for values above 25 there's an offset of 1. And for values above 87 there's an offset of 2. 115/118 and so on. Any idea why this might be? I would expect those two expressions to be identical.

My pipeline:

[
("vect", StemmedCountVectorizer(ngram_range=(1, 2), max_df=0.8, min_df=3),),
("tfidf", TfidfTransformer(use_idf=True)),
(
"clf",
SGDClassifier(
loss="modified_huber", penalty="l2", alpha=0.001, random_state=42
),
),
]


This seems to be a bug in the library.

https://github.com/scikit-learn/scikit-learn/issues/17035

The issue is predict_proba does not convert from its internal representation of categories to the dataset's representation before returning the data. Client-side we can convert using some internal APIs:

> [text_clf._final_estimator.classes_[max(enumerate(prob), key=lambda p: p[1])[0]] for prob in text_clf.predict_proba(test.data)]
[61, 16, 11, 16, 11, 89, 26, 33, 16, 57, 16, 118, 11, 26, ...]
> [*text_clf.predict(test.data)]
[61, 16, 11, 16, 11, 89, 26, 33, 16, 57, 16, 118, 11, 26, ...]


But the real fix needs to happen in the library.

• Not in fact a bug, but me misreading docs :). This is intended (though I still argue odd) – user95708 Jul 31 at 0:53