Validation: Gaussian noise in sub-topics

[ampudia19]

The idea is to estimate noise for every topic and sub-topic, allowing to

• identify differences in cluster granularity for otherwise same-labelled levels (ie. network Level 1 vs. semantic Level 1).
• estimate relative reductions in noise upon further splitting large clusters (very small ones are trivial).
• "Relative" gains of using a semantic approach vs. network (ie. if similar noise, little gain from semantic clustering)

In practical terms, this implies

• [ ] using semantic embeddings for the entities of each sub-topic and, assuming a Gaussian distribution, get the standard deviation.
• [ ] dimensionality reduction on the above to reduce overhead costs of multidimensional gaussians.

NB: This should correspond to bullet point "Relative difference in variance of embeddings / level" in Emily's comment on #12