Discrepancy in Scikit Learn Stratified Cross Validation

Question

I am seeing a discrepancy between classification performance between two cross validation technique using the same data. I was wondering if anyone can shed some light on this.

• Method 1: cross_validation.train_test_split
• Method 2: StratifiedKFold.

Two Examples with same data set

Data Set 5500[n_samples :: Class 1 = 500 ; Class 0 = 5000 ] by 193 Features

Method 1 [ Random Iteration with train_test_split ]

for i in range(0,5):
    X_tr, X_te, y_tr, y_te = cross_validation.train_test_split(X_train.values, y_train, test_size=0.2, random_state=i)
    clf = RandomForestClassifier(n_estimators=250, max_depth=None, min_samples_split=1, random_state=0, oob_score=True)
    y_score = clf.fit(X_tr, y_tr).predict(X_te)
    y_prob = clf.fit(X_tr, y_tr).predict_proba(X_te)
    cm = confusion_matrix(y_te, y_score)
    print cm
    fpr, tpr, thresholds = roc_curve(y_te,y_prob[:,1])
    roc_auc = auc(fpr, tpr);
    print "ROC AUC: ", roc_auc

Result of method 1

Iteration 1 ROC AUC:  0.91
[[998   4]
 [ 42  56]]

Iteration 5 ROC AUC:  0.88
[[1000    3]
 [  35   62]]

Method 2 [ StratifiedKFold cross validation ]

cv = StratifiedKFold(y_train, n_folds=5,random_state=None,shuffle=False)
clf = RandomForestClassifier(n_estimators=250, max_depth=None, min_samples_split=1, random_state=None, oob_score=True)
for train, test in cv:
    y_score = clf.fit(X_train.values[train], y_train[train]).predict(X_train.values[test])
    y_prob = clf.fit(X_train.values[train], y_train[train]).predict_proba(X_train.values[test])
    cm = confusion_matrix(y_train[test], y_score)
    print cm
    fpr, tpr, thresholds = roc_curve(y_train[test],y_prob[:,1])
    roc_auc = auc(fpr, tpr);
    print "ROC AUC: ", roc_auc

Result of method 2

Fold 1 ROC AUC:  0.76
Fold 1 Confusion Matrix
[[995   5]
 [ 92   8]]

Fold 5 ROC AUC:  0.77
Fold 5 Confusion Matrix
[[986  14]
 [ 76  23]]

Answer 1

train_test_split is not stratified. In the current development version and in the upcoming 0.17 you can do stratify=y to make it stratified, but not in 0.16.2. Also the random states are not fixed and used differently, so you can't expect the exact same outcome.