Discretizing

Converts continuous numerical data into discrete categories or intervals, mutually exclusive with encoding (encoder).

Usage Examples

Basic Usage

Preprocessor:
  demo:
    method: 'default'
    sequence:
      - missing
      - outlier
      - discretizing  # Use discretizing, not encoder

Custom K-bins Parameters

Preprocessor:
  custom:
    method: 'default'
    sequence:
      - missing
      - outlier
      - discretizing
    config:
      discretizing:
        age:
          method: 'discretizing_kbins'
          n_bins: 10                    # Divide into 10 bins
        income:
          method: 'discretizing_kbins'
          n_bins: 5                     # Divide into 5 bins

Available Processors

Processor	Description	Applicable Type	Output
`discretizing_kbins`	K-bins discretization	Numerical	Integer labels
`encoder_label`	Label encoding	Categorical	Integer labels

Note: In the discretizing sequence, categorical data will automatically use encoder_label for encoding.

Processor Details

discretizing_kbins

K-bins Discretization: Divides continuous values into k equal-width intervals.

Parameters:

n_bins (int, optional)
- Number of bins (k value)
- Default value: 5
- Example: n_bins: 10

Features:

Equal-width binning
Outputs integer labels (0, 1, 2, …)
Reduces data dimensionality and complexity

Example:

config:
  discretizing:
    age:
      method: 'discretizing_kbins'
      n_bins: 5

Discretization Example:

Original values: [18, 25, 35, 45, 55, 65]
n_bins = 5

Bin division:
[18-27.4) → 0
[27.4-36.8) → 1
[36.8-46.2) → 2
[46.2-55.6) → 3
[55.6-65] → 4

Discretized result: [0, 0, 1, 2, 3, 4]

Processing Logic

Numerical Data (discretizing_kbins)

Training phase (fit):
  Calculate and store bin boundaries

Transform phase (transform):
  Map values to integer labels based on bin boundaries

Inverse transform phase (inverse_transform):
  Restore integer labels to bin midpoint values

Categorical Data (encoder_label)

Training phase (fit):
  Create mapping from categories to integers

Transform phase (transform):
  Convert categories to integers based on mapping

Inverse transform phase (inverse_transform):
  Restore integers to categories based on mapping

Default Behavior

Default processing when using discretizing sequence:

Data Type	Default Processor	Description
Numerical	`discretizing_kbins`	K-bins discretization (k=5)
Categorical	`encoder_label`	Label encoding
Datetime	`discretizing_kbins`	K-bins discretization (k=5)

Differences from encoder

Feature	discretizing	encoder
Numerical Output	Discrete integers (0, 1, 2, …)	Continuous values or multiple columns
Use Case	Discretization needs	General encoding needs
Categorical Handling	Label encoding	Multiple encodings (Uniform/Label/OneHot)
With scaler	Usually not used	Usually used together
Sequence Position	Must be last step	Before scaler

Usage Restrictions

1. Mutually Exclusive with encoder

# ❌ Wrong: Cannot use both
Preprocessor:
  wrong:
    method: 'default'
    sequence:
      - missing
      - encoder       # Error!
      - discretizing  # Mutually exclusive with encoder

# ✅ Correct: Use only one
Preprocessor:
  correct:
    method: 'default'
    sequence:
      - missing
      - outlier
      - discretizing  # Correct

2. Must Be Last Step

# ❌ Wrong: Steps after discretizing
Preprocessor:
  wrong:
    method: 'default'
    sequence:
      - missing
      - discretizing
      - scaler        # Error! discretizing must be last

# ✅ Correct: discretizing is the last step
Preprocessor:
  correct:
    method: 'default'
    sequence:
      - missing
      - outlier
      - discretizing  # Correct: last step

Complete Example

Loader:
  load_data:
    filepath: 'data.csv'
    schema: 'schema.yaml'

Preprocessor:
  discretize_data:
    method: 'default'
    sequence:
      - missing
      - outlier
      - discretizing  # Note: No encoder and scaler
    config:
      # Missing value handling
      missing:
        age: 'missing_median'
        income: 'missing_mean'
      
      # Outlier handling
      outlier:
        age: 'outlier_iqr'
        income: 'outlier_iqr'
      
      # Discretizing configuration
      discretizing:
        # Numerical fields
        age:
          method: 'discretizing_kbins'
          n_bins: 10                    # Divide age into 10 bins
        income:
          method: 'discretizing_kbins'
          n_bins: 5                     # Divide income into 5 bins
        hours_per_week:
          method: 'discretizing_kbins'
          n_bins: 8                     # Divide hours into 8 bins
        
        # Categorical fields (automatically uses encoder_label)
        gender: 'encoder_label'
        education: 'encoder_label'

Synthesizer:
  synthesize:
    method: 'default'

Postprocessor:
  postprocess:
    method: 'default'

Use Cases

1. Simplify Data Distribution

# Simplify continuous distribution into discrete intervals
Preprocessor:
  simplify:
    method: 'default'
    sequence:
      - missing
      - discretizing
    config:
      discretizing:
        salary:
          method: 'discretizing_kbins'
          n_bins: 3  # Low/Medium/High three levels

2. Reduce Data Dimensionality

# Reduce numerical precision, decrease synthesis difficulty
Preprocessor:
  reduce_dimension:
    method: 'default'
    sequence:
      - missing
      - outlier
      - discretizing
    config:
      discretizing:
        age:
          method: 'discretizing_kbins'
          n_bins: 5
        score:
          method: 'discretizing_kbins'
          n_bins: 10

Notes

Mutual Exclusivity: Cannot be used with encoder simultaneously
Position Restriction: Must be the last step in the sequence
Restoration Precision: Restoration uses bin midpoints, won’t fully restore original values
Number of Bins: n_bins should be adjusted based on data range and requirements
Synthesizer Impact: Some synthesizers (like PAC-Synth, DPCTGAN) may produce floating-point numbers, which the system will automatically round
Use Case: Suitable for situations requiring discretized output, such as certain privacy protection algorithms
NA Handling: During post-processing, NA values will be removed before restoration