As an example, I have a dataset of available games.
Game A graphics presets are: Low, Medium, High, Ultra
Game B graphics presets are: Minimum, Balanced, Maximum
Game C graphics presets are: Ultra
Game A might not have the required feature scaling from Game B, therefore, using a global model (not a model per game) is essential.
Graphics preset is an ordinal categorical column.
Should a custom ordinal encoder encode features like:
Game A: 1, 2, 3, 4
Game B: 1, 2.5, 4
Game C: 2.5
or
Game A: 1, 2, 3, 4
Game B: 1, 2, 3
Game C: 1
Later I would standard scale all features (except one-hot game A, B, C, ... category) to predict a linear value.