Uniform Encoder

When converting categorical variables to continuous values, the Uniform Encoder can provide better performance for generative models. This method, proposed by datacebo, maps each category to a specific range in a uniform distribution, with the range size determined by the relative proportion of each category in the data.

Principles

Core Concepts

• Map discrete categories to [0,1] interval
• Determine mapping ranges based on category frequencies
• Generate random values within ranges as encoded results

Advantages

1. Converts discrete distributions to continuous, facilitating modeling
2. Fixed value range [0,1], simplifying category restoration
3. Preserves original distribution information, with more frequent categories having larger sampling probabilities

Implementation Example

Consider a categorical variable with three categories ‘a’, ‘b’, ‘c’, appearing in proportions 1:3:1:

mapping = {
   'a': [0.0, 0.2),  # 20% range
   'b': [0.2, 0.8),  # 60% range
   'c': [0.8, 1.0]   # 20% range
}

Encoding Process

• Category ‘a’ → Random value in [0.0, 0.2)
• Category ‘b’ → Random value in [0.2, 0.8)
• Category ‘c’ → Random value in [0.8, 1.0]

Restoration Process

• Check which range contains the value
• Map back to corresponding category

Usage Guidelines

• Best suited for features with fewer categories
• Particularly effective for imbalanced category distributions
• Can be combined with other preprocessing methods (e.g., scaling)

References

• Improving Synthetic Data Quality with the Uniform Encoder

Appx.: Available Evaluation Methods

Available Evaluation Methods

The evaluator supports three main types of evaluation methods:

• Privacy Risk Assessment evaluates the level of privacy protection in synthetic data. Including:

  • Singling Out Risk: Assesses the ability to identify specific individuals in the data
  • Linkability Risk: Assesses the ability to link records of the same individual across different datasets
  • Inference Risk: Assesses the ability to infer attributes from known information

• Data Fidelity Assessment evaluates the fidelity of synthetic data. Including:

  • Diagnostic Report: Examines data structure and basic properties
  • Quality Report: Evaluates statistical distribution similarity

• Data Utility Assessment evaluates the practical value of synthetic data. Including:

  • Classification Utility: Compares classification model performance
  • Regression Utility: Compares regression model performance
  • Clustering Utility: Compares clustering results

Privacy Risk Assessment

Singling Out Risk Assessment

Evaluates the ability to identify records of specific individuals in the data. Results are scored from 0 to 1, where higher numbers indicate greater privacy risk.

Parameters

• n_attacks (int, default=2000): Number of attack attempts (unique queries)
• n_cols (int, default=3): Number of columns used per query
• max_attempts (int, default=500000): Maximum attempts to find successful attacks

Returns

• pd.DataFrame: Evaluation results dataframe containing:
  • risk: Privacy risk score (0-1)
  • risk_CI_btm: Privacy risk confidence interval lower bound
  • risk_CI_top: Privacy risk confidence interval upper bound
  • attack_rate: Main privacy attack success rate
  • attack_rate_err: Main privacy attack success rate error
  • baseline_rate: Baseline privacy attack success rate
  • baseline_rate_err: Baseline privacy attack success rate error
  • control_rate: Control group privacy attack success rate
  • control_rate_err: Control group privacy attack success rate error

Linkability Risk Assessment

Evaluates the ability to link records belonging to the same individual across different datasets. Results are scored from 0 to 1, where higher numbers indicate greater privacy risk.

Parameters

• n_attacks (int, default=2000): Number of attack attempts
• max_n_attacks (bool, default=False): Whether to force maximum number of attacks
• aux_cols (Tuple[List[str], List[str]]): Auxiliary information columns, for example:

  aux_cols = [
      ['gender', 'zip_code'],  # public data
      ['age', 'medical_history']  # private data
  ]
• n_neighbors (int, default=10): Number of nearest neighbors to consider

Returns

• pd.DataFrame: Evaluation results dataframe in the same format as singling out risk assessment

Inference Risk Assessment

Evaluates the ability to infer attributes from known information. Results are scored from 0 to 1, where higher numbers indicate greater privacy risk.

Parameters

• n_attacks (int, default=2000): Number of attack attempts
• max_n_attacks (bool, default=False): Whether to force maximum number of attacks
• secret (str): Column to be inferred
• aux_cols (List[str], optional): Columns used for inference, defaults to all columns except secret

Returns

• pd.DataFrame: Evaluation results dataframe in the same format as singling out risk assessment

Data Fidelity Assessment

Diagnostic Report

Validates synthetic data structure and basic properties.

Parameters

None

Returns

• pd.DataFrame: Evaluation results dataframe containing:
  • Score: Overall diagnostic score
  • Data Validity: Data validity score
    • KeyUniqueness: Primary key uniqueness
    • BoundaryAdherence: Numeric range compliance
    • CategoryAdherence: Category compliance
  • Data Structure: Data structure score
    • Column Existence: Column presence
    • Column Type: Column type compliance

Quality Report

Evaluates statistical similarity between original and synthetic data.

Parameters

None

Returns

• pd.DataFrame: Evaluation results dataframe containing:
  • Score: Overall quality score
  • Column Shapes: Column distribution similarity
    • KSComplement: Continuous variable distribution similarity
    • TVComplement: Categorical variable distribution similarity
  • Column Pair Trends: Column relationship preservation
    • Correlation Similarity: Correlation preservation
    • Contingency Similarity: Contingency table similarity

Data Utility Assessment

Classification Utility Assessment

Compares classification model performance between original and synthetic data using logistic regression, support vector machine, random forest, and gradient boosting (all with default parameters).

Parameters

• target (str): Classification target column

Returns

• pd.DataFrame: Evaluation results dataframe containing:
  • ori_mean: Original data model average F1 score
  • ori_std: Original data model F1 score standard deviation
  • syn_mean: Synthetic data model average F1 score
  • syn_std: Synthetic data model F1 score standard deviation
  • diff: Synthetic data improvement over original data

Regression Utility Assessment

Compares regression model performance between original and synthetic data using linear regression, random forest regression, and gradient boosting regression (all with default parameters).

Parameters

• target (str): Prediction target column (numeric)

Returns

• pd.DataFrame: Evaluation results dataframe containing:
  • ori_mean: Original data model average R² score
  • ori_std: Original data model R² score standard deviation
  • syn_mean: Synthetic data model average R² score
  • syn_std: Synthetic data model R² score standard deviation
  • diff: Synthetic data improvement over original data

Clustering Utility Assessment

Compares K-means clustering algorithm (with default parameters) results between original and synthetic data.

Parameters

• n_clusters (list, default=[4, 5, 6]): List of cluster counts

Returns

• pd.DataFrame: Evaluation results dataframe containing:
  • ori_mean: Original data average silhouette score
  • ori_std: Original data silhouette score standard deviation
  • syn_mean: Synthetic data average silhouette score
  • syn_std: Synthetic data silhouette score standard deviation
  • diff: Synthetic data improvement over original data