Jinyi-Liu

src/scala/com/twitter/interaction_graph/bqe/training/labeled_candidates.sql In this case, the rows with negative labels are ignored so the labels being trained are {0,1}. Why don't we include the negative label and use {-1,0,1}?