TABLE 3.

The error rates for the 50 posts most likely to be labeled correctly for each of the learning methods

Trial:	pos_neg (classifier trained on all 53 labeled posts)	pos_all (classifier trained on the 35 “positive” labeled posts and the 4,759 unlabeled posts, treating the unlabeled posts as if they were labeled “negative”)	pu_learning (classifier trained on 80% of the 35 “positive” labeled posts, and the 4,759 unlabeled posts, treating the unlabeled posts as if they were labeled “negative”)	wmd_1 (classifier using the word mover’s distance to the single closest “positive” post)	wmd_5 (classifier using the average of the word mover’s distance to the five closest “positive” posts)
Top-50 error rate:	4%	12%	8%	12%	4%