Daemony

# Problem For single label classification, hitrate at top k is useful, more than NAUC and acc.

# Problem Text domain, **prediction error** when class_num (500~1000) and train_num(~400000) is large.

# Question In the tutorial and examples, it should support *stratified splitting train/test* processing, more than random split. # Description If *class_num* is big, or imbalanced problem is serious, the...

# Question For text data, ** language detect make errors ** when meeting traditional chinese and uppercase letters. # Description > t=我有一个梦想, lang=zh-cn t=I have a dream, lang=en t=我有一個夢想, lang=ko...