Compare Drain-like approaches extracting log keys and log key parameters to the embedding approach.

[dhonza]

Take all appropriate loglizer AD methods and make them work with the fastText (and possibly other embeddings) embeddings instead of log key categories.

Possible problem: log embedding dimension might be too high for methods like Isolation Forests.

[dhonza]

@korytmar does a lot of these experiments in his Master Thesis (to be submitted by the end of May).