Empirical study of the value of weak supervision

This code allows the reader to replicate all the experiments carried out in our study, and obtain the figures.

• figA relates the size of the candidate sets with the proportion of weakly labeled examples.

• figB relates the number of fully labeled samples with the proportion of weakly labeled ones.

• figC relates the probability of co-occurrence with other parameters (size of candidate sets, proportion of weakly labeled samples).

• onlyfull refers to baseline experiments where only the fully labeled subsets is used to learn the model.

• real_model refers to experiments that use the real generative model to establish the maximum reachable.

  • With real data, this real model is simulated.

• All the models and datasets are generated once using gen_models_and_data scripts.

The two notebooks collect experimental results and make the plots for synthetic and real data.