Model random seed management
Background
Some Machine Learning models offer to set the random seed as input.
Default random seed are constant values at the moment.
As a user of Polaris I would like to set how randomness is managed in these models all together (for now). For auditing purposes, it is useful to know how randomness has been set.
The Call
What this issue would solve are the following:
- choice of the default behaviour of polaris: constant random seeds, or random random-seeds, or something else
- implementation of a new configuration keyword where the user can chose a random mode:
- constant seed, preset
- random (like -1 could mean random, if we don't want to add a
random_mode
variable :) ) - any other option (such as
unset
which fall for default of machine learning libraries)
- Implementation in both feature extraction+selection and in cross dependency analysis.
Side notes
Label cli
used to mean it has impact on keywords usable in user-defined configuration.