MedGANDataType

Enums · AiDotNet.Enums

The kind of variables medGAN’s autoencoder is being fitted to. This selects the autoencoder’s activations and its pre-training reconstruction loss.

For Beginners

Patient records come in two shapes. Either a column says “does this patient have diagnosis X?” — a yes/no, which is Binary — or it says “how many times did this happen?” — a count, which is Count. The two need different maths, so you tell the model which one you have. MixedTabular is for the common real-world case where a single table has some of each plus ordinary numeric columns.

How It Works

medGAN (Choi et al., arXiv:1703.06490) is defined for “high-dimensional discrete variables (e.g., binary and count features)” and gives a separate activation/loss pair for each. §3.2 states the count case as the mean squared loss (Eq. 2) and the binary case as the cross entropy loss (Eq. 3).

Fields

Field Summary
MedGANDataType Binary Binary variables.
MedGANDataType Count Count variables.
MedGANDataType MixedTabular Mixed tabular data as produced by TabularDataTransformer: each continuous column becomes one mode-normalized scalar plus a one-hot mode indicator, and each categorical column becomes a one-hot group.
int value__