Logistic Regression vs. Linear SVM over 10 bootstrap samples.
Log. Reg. trained for 30 epochs with 0.1 learning rate, 0.5 momentum, no regularization.
Linear SVM trained using SVMlight, selecting best C in [1e-5, 5e-5, 1e-4, ..., 1e+9]. Tested on out-of-sample examples for each bootstrap.
Negatives: N2 (Direct annotations to any ancestor node are not used as negatives).

AUCs: Mean ± StDev of Area under ROC Curve over 10 bootstrap samples.
C.Median: AUC of Bagging classifier, aggregating predictions by median (voting).
(Each Bagging weak classifier only contributes test output on examples held-out from it.)
Pos: Number of positive examples available.
Recall ≥ 10%: Best precision over recalls ≥0.10 and its highest corresponding recall.
TP ≥ 10%: Best precision over recalls ≥(10/Pos) and its highest corresponding recall.

Click column headers to sort! Click again to sort descending!