Open justin13601 opened 3 months ago
We hope to directly support tasks whose labels cannot be reliably summarized as a single numerical value (ie., a sequential decoding task where we want to predict a sequence of outputs, such as future treatment sequences). For tasks that rely on time-to-event as a paradigm, labels can currently be manually extracted via the window start
/end
times (ie., subtracting start
from end
of your window containing your event of interest, and then bucketing appropriately post-hoc).
To-do: merge this and #45
Right now we basically only support binary classification. This is ok for now, but we could aim to support more complex tasks:
Note this overlaps with #72