Currently, the models are just thrown out when a drift detector alarm is raised (to be reinitialized by new samples which may come after). A better approach might be to start buffering samples since a warning is raised and use them to immediately reinitialize the model after an alarm.
Currently, the models are just thrown out when a drift detector alarm is raised (to be reinitialized by new samples which may come after). A better approach might be to start buffering samples since a warning is raised and use them to immediately reinitialize the model after an alarm.