Code/strategy to add active learning classifications to training set

[bfhealy]

Scope code currently downloads and sums votes for each classification in a group of sources, but we do not yet have the code or strategy to grow the existing training set from that information. One potential strategy is:

• Add all sources marked as 'Labelled' to the existing training set
• Remove a classification if its votes sum to < 0
• Keep a classification if its votes sum to >= 0
• In cases of duplicate classifications, assign the probability from the one most recently posted

One problem with the above strategy is that there is ambiguity about the reliability of classifications with a net vote sum of 0. Perhaps no labeler was sure about that classification, or an even number disagreed - should it be kept or not?

Perhaps each newly added portion of the training set learning should be uploaded to a unique group on Fritz. This would serve as a kind of 'version control' for this part of the project.