Custom quantization borders and missing value modes

Contains

Custom quantization borders and the method for processing missing values for the dataset numerical features.

Format

  • Each line contains information regarding a single border and optionally the missing values mode settings for the corresponding feature.

  • Different missing values mode settings can not be specified for a single feature. The value either has to be set only on one line or should be the same on different lines that contain information regarding a single feature.

  • Supported missing values modes:

    • "Forbidden" — Missing values are not supported, their presence is interpreted as an error.
    • "Min" — Missing values are processed as the minimum value (less than all other values) for the feature. It is guaranteed that a split that separates missing values from all other values is considered when selecting trees.
    • "Max" — Missing values are processed as the maximum value (greater than all other values) for the feature. It is guaranteed that a split that separates missing values from all other values is considered when selecting trees.
  • The missing values mode for the feature is defined by the value of the --nan-mode (nan_mode) training parameter if not specified in this file.

  • Format of a single line:

    <zero-based feature ID><\t><border value><\t><missing values mode (optional)>
    

Example

The following description contains two borders for features indexed 0 and 2 and missing values settings for each of these features:

0<\t>0.25<\t>Forbidden
0<\t>0.75<\t>Forbidden
2<\t>0.3
2<\t>0.85<\t>Max