meshulam
commented
7 years ago After conversation today:
Article
- has one: LearnedData
- has many: EnteredData
- reviewed boolean (just whether an EnteredData entry exists?)
LearnedData
- probabilities for each category
- relevance probability
- overall confidence 0-1 (calculated from above?)
- model info string (version, date, hyperparameters)
EnteredData
- user ID
- boolean for each category
- relevant/irrelevant
- date
Article is responsible for figuring out the 'definitive' classification based on what data is associated with it
When we run the classification algo on the full dataset, we probably don't want to modify the article records directly. Especially if human coders have already classified it.
Idea: derived metadata (categories, relevant/not, location) should live in its own related table, which includes fields for how it was derived (human or algo). Could there be multiple classifications for an article, and choose which one to use?