Naive Bayes:

Semi-supervised:

• use integers
• when use unlabeled data, we can make the new label fraction
• choose top-k