The Options class of the ranking trainers (FastTree and LightGbm) defaults to RowGroupColumnName = null.
This is:
- Inconsistent with the simple constructor where
RowGroupColumnName defaults to GroupId
- Not desirable as in ranking the row group is very important for correct training
Here are the lines where the row group column name is set:
|
/// <summary> |
|
/// Column to use for example groupId. |
|
/// </summary> |
|
[Argument(ArgumentType.AtMostOnce, HelpText = "Column to use for example groupId", ShortName = "groupId", SortOrder = 5, Visibility = ArgumentAttribute.VisibilityType.EntryPointsOnly)] |
|
public string RowGroupColumnName = null; |
We need to update the default and align it with the simple constructor.
The
Optionsclass of the ranking trainers (FastTreeandLightGbm) defaults toRowGroupColumnName = null.This is:
RowGroupColumnNamedefaults toGroupIdHere are the lines where the row group column name is set:
machinelearning/src/Microsoft.ML.Data/Training/TrainerInputBase.cs
Lines 106 to 110 in 738e5d5
We need to update the default and align it with the simple constructor.