-After the 1st phase matrix factorization, compute a new matrix factorization with the harassment/abuse tag as the label, instead of the overall helpful rating like normal.
-Add a large rater reputation penalty on raters who have rated any notes with extremely high harassment/abuse scores as helpful
-Support BCEWithLogits loss and pos_weight in order to train the imbalanced binary matrix factorization well
-After the 1st phase matrix factorization, compute a new matrix factorization with the harassment/abuse tag as the label, instead of the overall helpful rating like normal. -Add a large rater reputation penalty on raters who have rated any notes with extremely high harassment/abuse scores as helpful -Support BCEWithLogits loss and pos_weight in order to train the imbalanced binary matrix factorization well