python - Using the predict_proba() function of RandomForestClassifier in the safe and right way -

i'm using scikit-learn apply machine learning algorithm on datasets. need have probabilities of labels/classes instated of labels/classes themselves. instead of having spam/not spam labels of emails, wish have example: 0.78 probability given email spam.

for such purpose, i'm using predict_proba() randomforestclassifier following:

clf = randomforestclassifier(n_estimators=10, max_depth=none,     min_samples_split=1, random_state=0) scores = cross_val_score(clf, x, y) print(scores.mean())  classifier = clf.fit(x,y) predictions = classifier.predict_proba(xtest) print(predictions) 

and got results:

 [ 0.4  0.6]  [ 0.1  0.9]  [ 0.2  0.8]  [ 0.7  0.3]  [ 0.3  0.7]  [ 0.3  0.7]  [ 0.7  0.3]  [ 0.4  0.6] 

where second column class: spam. however, have 2 main issues results not confident. first issue results represent probabilities of labels without being affected size of data? second issue results show 1 digit not specific in cases 0.701 probability different 0.708. there way next 5 digit example?

many in advance time in reading these 2 issues , questions.

1. i more 1 digit in results, sure not due dataset ? (for example using small dataset yield simple decision trees , 'simple' probabilities). otherwise may display shows 1 digit, try print predictions[0,0].

2. i not sure understand mean "the probabilities aren't affected size of data". if concern don't want predict, eg, many spams, done use threshold t such predict 1 if proba(label==1) > t. way can use threshold balance predictions, example limit global probabilty of spams. , if want globally analyse model, compute area under curve (auc) of receiver operating characteristic (roc) curve (see wikipedia article here). roc curve description of predictions depending on threshold t.

hope helps!