There are a few query strategies available, and depending on the needs of the simulation/review some will work better than others.
Parameters should be under the section
As it says: randomly select samples with no regard to model assigned probabilities.
Choose the most uncertain samples according to the model (i.e. closest to 0.5 probability). Probably doesn’t work very well in the case of LSTM’s, since the probabilities are rather arbitrary.
Choose the most likely samples to be included according to the model.
Use a combination of random and max sampling. By default it does 5% random sampling and 95% max sampling. Works well in combination with the triple balance strategy. This parameter can be set in the configuration file:
# Set to 5% random, 95% max sampling. rand_max_frac=0.05