give lower weights to samples with weak results for model training

[aliinreallife]

Intro

Currently we use weak supervised learning provided by snorkel to replace weak acc/rej with acc/rej to have a binary classification. We can try to give lower weights to weak results and train our model with them to see if we would have a better model.

tasks

• prior work research
• check the results
• implement

[Nikronic]

@aliinreallife Please check #10 and #41. I think these can be helpful (no idea).

A blog related to positive unlabeled learning (sort of replacement for snorkel): https://roywrightme.wordpress.com/2017/11/16/positive-unlabeled-learning/