Package {rsdv}

rsdv: Synthetic Tabular Data Generation with Gaussian Copulas

Description

rsdv generates synthetic tabular data from real datasets using Gaussian copula models, with parametric marginal selection for numerical columns and a cumulative-frequency embedding that brings categorical and boolean columns into the same joint copula. It includes a metadata system, declarative constraints, conditional sampling, and quality, validity, and privacy reports modeled on those of the Python SDMetrics library.

Details

The main entry points are:

• metadata() — describe column types and primary keys.

• gaussian_copula_synthesizer() + fit() + sample() — fit a synthesizer and generate rows.

• sample_conditions() — generate rows that hold given categorical or boolean values fixed.

• quality_report(), diagnostic_report(), privacy_report() — evaluate the synthetic data.

Author(s)

Maintainer: Kailas Venkitasubramanian kailasv@gmail.com

See Also

Useful links:

• https://kvenkita.github.io/rsdv/

• https://github.com/kvenkita/rsdv

• Report bugs at https://github.com/kvenkita/rsdv/issues

Add a constraint to metadata

Description

Add a constraint to metadata

Usage

add_constraint(meta, constraint)

Arguments

Value

Updated rsdv_metadata (for piping).

Examples

meta <- metadata() |>
  set_column_type("low", "numerical") |>
  set_column_type("high", "numerical") |>
  add_constraint(inequality_constraint("low", "high", type = "lt"))

Adult Income dataset (500-row sample)

Description

A 500-row random sample of the UCI Adult Income dataset, used in package examples and vignettes.

Usage

adult_income

Format

A tibble with 500 rows and 16 variables:

id

Row identifier (integer)

age

Age in years (integer)

workclass

Employment type (character)

fnlwgt

Final weight, a census sampling weight (integer)

education

Highest level of education (character)

education_num

Education encoded as an integer (integer)

marital_status

Marital status (character)

occupation

Occupation category (character)

relationship

Relationship to householder (character)

race

Race (character)

sex

Sex (character)

capital_gain

Capital gains (integer)

capital_loss

Capital losses (integer)

hours_per_week

Hours worked per week (integer)

native_country

Country of origin (character)

income

Income bracket: ⁠<=50K⁠ or ⁠>50K⁠ (character)

Source

https://archive.ics.uci.edu/dataset/2/adult

Attribute disclosure risk

Description

Estimates the fraction of synthetic rows where a sensitive column value can be correctly inferred from known columns via a k-NN lookup in the real training data.

Usage

attribute_disclosure_risk(real, synthetic, sensitive_col, known_cols, k = 1L)

Arguments

Value

A scalar in [0, 1]; lower = more private.

Examples

real <- data.frame(age = sample(20:60, 50, replace = TRUE),
                   income = sample(c("low", "high"), 50, replace = TRUE),
                   stringsAsFactors = FALSE)
syn  <- real[sample(50), ]
attribute_disclosure_risk(real, syn, sensitive_col = "income", known_cols = "age")

Plot a diagnostic report

Description

Bar chart of per-column validity scores.

Usage

## S3 method for class 'rsdv_diagnostic_report'
autoplot(object, ...)

Arguments

Value

A ggplot object.

Examples

meta  <- metadata(adult_income)
syn   <- gaussian_copula_synthesizer(meta) |> fit(adult_income)
synth <- sample(syn, n = 500)
ggplot2::autoplot(diagnostic_report(adult_income, synth, meta))

Plot a privacy report

Description

Plots the NNDR score as a gauge-style bar.

Usage

## S3 method for class 'rsdv_privacy_report'
autoplot(object, ...)

Arguments

Value

A ggplot object.

Examples

syn <- gaussian_copula_synthesizer(metadata(adult_income)) |> fit(adult_income)
synth <- sample(syn, n = 500)
pr <- privacy_report(adult_income, synth)
ggplot2::autoplot(pr)

Plot a quality report

Description

Produces a bar chart of per-column similarity scores, with a horizontal line at the overall score.

Usage

## S3 method for class 'rsdv_quality_report'
autoplot(object, ...)

Arguments

Value

A ggplot object.

Examples

syn <- gaussian_copula_synthesizer(metadata(adult_income)) |> fit(adult_income)
synth <- sample(syn, n = 500)
qr <- quality_report(adult_income, synth, metadata(adult_income))
ggplot2::autoplot(qr)

Check a single constraint against each row of a data frame

Description

Check a single constraint against each row of a data frame

Usage

check_constraint(data, constraint)

Arguments

Value

Logical vector of length nrow(data).

Examples

df <- data.frame(a = c(1, 2, 3), b = c(1, 2, 9))
check_constraint(df, equality_constraint("a", "b"))

Check all constraints in metadata against a data frame

Description

Check all constraints in metadata against a data frame

Usage

check_constraints(data, meta)

Arguments

Value

Logical vector of length nrow(data). TRUE = row passes all constraints.

Examples

meta <- metadata() |>
  set_column_type("x", "numerical") |>
  add_constraint(custom_constraint(function(row) row$x > 0))
check_constraints(data.frame(x = c(1, -1, 2)), meta)

Contingency similarity between real and synthetic categorical column pairs

Description

For each pair of categorical columns, compares the joint (normalized contingency) distributions of real and synthetic data via total variation distance, scoring 1 - TVD (the SDMetrics ContingencySimilarity score). This is the categorical analogue of correlation similarity and captures categorical-vs-categorical dependence.

Usage

contingency_similarity(real, synthetic, meta)

Arguments

Value

A list with pairs (a tibble of column_1, column_2, score) and score (the mean over pairs; 1 when there are fewer than two categorical columns).

Examples

meta  <- metadata(adult_income)
syn   <- gaussian_copula_synthesizer(meta) |> fit(adult_income)
synth <- sample(syn, n = 500)
contingency_similarity(adult_income, synth, meta)

Correlation similarity between real and synthetic numerical column pairs

Description

For each pair of numerical columns, computes ⁠1 - |corr_real - corr_syn| / 2⁠ (the SDMetrics CorrelationSimilarity score), where corr is the Pearson correlation. Returns one row per pair plus the mean.

Usage

correlation_similarity(real, synthetic, meta)

Arguments

Value

A list with pairs (a tibble of column_1, column_2, score) and score (the mean over pairs; 1 when there are fewer than two numerical columns).

Examples

syn       <- gaussian_copula_synthesizer(metadata(adult_income)) |> fit(adult_income)
synth_data <- sample(syn, n = 500)
correlation_similarity(adult_income, synth_data, metadata(adult_income))

Constraint: arbitrary row-wise predicate

Description

Constraint: arbitrary row-wise predicate

Usage

custom_constraint(fn)

Arguments

Value

An rsdv_constraint object.

Examples

custom_constraint(function(row) row$x > 0)

Generate a diagnostic (validity) report for synthetic data

Description

Checks whether synthetic data is structurally valid against the real data and metadata — independent of how closely it matches the real distributions (that is the job of quality_report()). Mirrors the SDMetrics DiagnosticReport two-property hierarchy:

Usage

diagnostic_report(real, synthetic, metadata)

Arguments

Details

• Data Validity — per-column checks:

  • numerical: boundary adherence (fraction of values within the real min/max range),

  • categorical: category adherence (fraction of values whose category was seen in the real data),

  • boolean: always valid,

  • primary key: key uniqueness (all values unique and non-missing).

• Data Structure — fraction of expected columns present in the synthetic data.

Missing (NA) values are excluded from adherence denominators, since missingness is modeled separately.

Value

An rsdv_diagnostic_report object.

Examples

meta  <- metadata(adult_income)
syn   <- gaussian_copula_synthesizer(meta) |> fit(adult_income)
synth <- sample(syn, n = 500)
diagnostic_report(adult_income, synth, meta)

Constraint: two columns must be equal row-wise

Description

Constraint: two columns must be equal row-wise

Usage

equality_constraint(col_a, col_b)

Arguments

Value

An rsdv_constraint object.

Examples

equality_constraint("city", "city_copy")

Constraint: only observed column combinations are valid

Description

Constraint: only observed column combinations are valid

Usage

fixed_combinations_constraint(columns, reference_data)

Arguments

Value

An rsdv_constraint object.

Examples

ref <- data.frame(city = c("NY", "LA"), state = c("NY", "CA"),
                  stringsAsFactors = FALSE)
fixed_combinations_constraint(c("city", "state"), ref)

Create a Gaussian Copula synthesizer

Description

Fits a single Gaussian copula over all modeled columns. Numerical columns use a fitted parametric marginal (see default_distribution); categorical and boolean columns are embedded into the copula via their cumulative-frequency intervals, so cross-column dependence (numeric vs. categorical, categorical vs. categorical) is preserved.

Usage

gaussian_copula_synthesizer(
  metadata,
  enforce_min_max = TRUE,
  numerical_distributions = list(),
  default_distribution = "auto"
)

Arguments

Value

An unfitted gaussian_copula_synthesizer object.

Examples

meta <- metadata(adult_income) |>
  set_column_type("age", "numerical") |>
  set_column_type("occupation", "categorical")
syn <- gaussian_copula_synthesizer(meta, default_distribution = "auto")
syn <- fit(syn, adult_income)

Constraint: col_a must be less than / greater than col_b

Description

Constraint: col_a must be less than / greater than col_b

Usage

inequality_constraint(col_a, col_b, type = c("lt", "lte", "gt", "gte"))

Arguments

Value

An rsdv_constraint object.

Examples

inequality_constraint("low", "high", type = "lt")

Check whether a synthesizer has been fitted

Description

Check whether a synthesizer has been fitted

Usage

is_fitted(x)

Arguments

Value

TRUE if fit() has been called; FALSE otherwise.

Examples

syn <- gaussian_copula_synthesizer(metadata())
is_fitted(syn)  # FALSE before fitting

Kolmogorov-Smirnov similarity score per numerical column

Description

Kolmogorov-Smirnov similarity score per numerical column

Usage

ks_similarity(real, synthetic, meta)

Arguments

Value

A tibble with columns column (chr) and score (dbl, 0–1, higher = better).

Examples

syn   <- gaussian_copula_synthesizer(metadata(adult_income)) |> fit(adult_income)
synth <- sample(syn, n = 500)
ks_similarity(adult_income, synth, metadata(adult_income))

Load metadata from a JSON file

Description

Load metadata from a JSON file

Usage

load_metadata(path)

Arguments

Value

An rsdv_metadata object.

Examples

meta <- metadata() |> set_column_type("age", "numerical")
tmp  <- tempfile(fileext = ".json")
save_metadata(meta, tmp)
load_metadata(tmp)

Create a metadata object describing a dataset's column types

Description

Create a metadata object describing a dataset's column types

Usage

metadata(data = NULL)

Arguments

Value

An rsdv_metadata object.

Examples

meta <- metadata(adult_income) |>
  set_column_type("age", "numerical") |>
  set_column_type("occupation", "categorical")

Deserialize metadata from a JSON string

Description

Deserialize metadata from a JSON string

Usage

metadata_from_json(json)

Arguments

Value

An rsdv_metadata object.

Examples

json <- metadata_to_json(metadata() |> set_column_type("age", "numerical"))
metadata_from_json(json)

Serialize metadata to a JSON string

Description

Serialize metadata to a JSON string

Usage

metadata_to_json(meta)

Arguments

Value

A JSON character string.

Examples

meta <- metadata() |> set_column_type("age", "numerical")
json <- metadata_to_json(meta)
meta2 <- metadata_from_json(json)

ML efficacy: train-on-synthetic / test-on-real accuracy ratio (TSTR)

Description

Trains an rpart decision tree on synthetic data and on a real training split, evaluates both on a real held-out test set, and returns the ratio TSTR / TRTR. A score near 1 means synthetic data is as informative as real data for this prediction task.

Usage

ml_efficacy(real, synthetic, meta, target_col, test_fraction = 0.2)

Arguments

Value

A list with elements tstr (accuracy), trtr (accuracy), and score (ratio, capped at 1).

Examples

meta      <- metadata(adult_income)
syn       <- gaussian_copula_synthesizer(meta) |> fit(adult_income)
synth_data <- sample(syn, n = 500)
ml_efficacy(adult_income, synth_data, meta, target_col = "income")

Nearest-Neighbor Distance Ratio privacy score

Description

For each synthetic row, computes the ratio of its distance to the nearest real row vs. its distance to the second-nearest real row. A high ratio (close to 1) means the synthetic row is not unusually close to any specific real row — low disclosure risk. Score = mean(ratio > 0.5).

Usage

nndr(real, synthetic)

Arguments

Value

A scalar score in [0, 1]; higher = more private.

Examples

real <- data.frame(x = rnorm(50), y = rnorm(50))
syn  <- data.frame(x = rnorm(50), y = rnorm(50))
nndr(real, syn)

Print method for rsdv_diagnostic_report

Description

Print method for rsdv_diagnostic_report

Usage

## S3 method for class 'rsdv_diagnostic_report'
print(x, ...)

Arguments

Value

x, invisibly.

Print method for rsdv_metadata

Description

Print method for rsdv_metadata

Usage

## S3 method for class 'rsdv_metadata'
print(x, ...)

Arguments

Value

x, invisibly.

Examples

print(metadata())

Print method for rsdv_privacy_report

Description

Print method for rsdv_privacy_report

Usage

## S3 method for class 'rsdv_privacy_report'
print(x, ...)

Arguments

Value

x, invisibly.

Examples

syn <- gaussian_copula_synthesizer(metadata(adult_income)) |> fit(adult_income)
synth <- sample(syn, n = 500)
pr <- privacy_report(adult_income, synth)
print(pr)

Print method for rsdv_quality_report

Description

Print method for rsdv_quality_report

Usage

## S3 method for class 'rsdv_quality_report'
print(x, ...)

Arguments

Value

x, invisibly.

Generate a privacy report comparing real and synthetic data

Description

Generate a privacy report comparing real and synthetic data

Usage

privacy_report(real, synthetic, sensitive_col = NULL, known_cols = NULL)

Arguments

Value

An rsdv_privacy_report object.

Examples

syn   <- gaussian_copula_synthesizer(metadata(adult_income)) |>
  fit(adult_income)
synth <- sample(syn, n = 500)
pr    <- privacy_report(adult_income, synth)
print(pr)

Generate a quality report comparing real and synthetic data

Description

Aggregates metrics into the two-property hierarchy used by SDMetrics:

Usage

quality_report(real, synthetic, metadata, target_col = NULL)

Arguments

Details

• Column Shapes — per-column marginal fidelity: KS similarity for numerical columns and TVD similarity for categorical columns.

• Column Pair Trends — pairwise dependence: correlation similarity for numerical pairs and contingency similarity for categorical pairs.

The overall score is the mean of the two property scores, so a table with many categorical columns and few numerical ones is not weighted by raw column counts. ML efficacy, when requested, is reported separately and does not enter the overall score (matching SDMetrics).

Value

An rsdv_quality_report object.

Examples

meta  <- metadata(adult_income) |>
  set_column_type("age", "numerical") |>
  set_column_type("occupation", "categorical")
syn   <- gaussian_copula_synthesizer(meta) |> fit(adult_income)
synth <- sample(syn, n = 500)
qr    <- quality_report(adult_income, synth, meta)
print(qr)

Objects exported from other packages

Description

These objects are imported from other packages. Follow the links below to see their documentation.

generics

fit

Sample synthetic rows from a fitted synthesizer

Description

Dispatches to the synthesizer-specific method when x is an rsdv_synthesizer. For plain R vectors, integers, or characters it falls back to base::sample(), preserving backward compatibility.

Usage

sample(x, n = NULL, ...)

Arguments

Value

When x inherits from rsdv_synthesizer, a data frame of n synthetic rows whose columns match the metadata. When x is any other object, the value returned by base::sample() — typically a vector of the same type as x and length n.

Examples

# Falls back to base::sample for non-synthesizer objects:
sample(1:10, 3)


meta  <- metadata(adult_income) |>
  set_column_type("age",    "numerical") |>
  set_column_type("income", "categorical")
syn   <- gaussian_copula_synthesizer(meta) |> fit(adult_income)
synth <- sample(syn, n = 100)
head(synth)

Sample synthetic rows that match fixed column values (conditional sampling)

Description

Generates rows in which one or more categorical or boolean columns are held to specified values, via rejection sampling against the fitted copula. This preserves the modeled dependence between the conditioned columns and the rest of the table (unlike overwriting values after the fact).

Usage

sample_conditions(x, conditions, max_tries = 100L)

Arguments

Value

A data frame of synthetic rows satisfying the conditions.

Examples

meta <- metadata(adult_income)
syn  <- gaussian_copula_synthesizer(meta) |> fit(adult_income)
sample_conditions(syn, data.frame(income = ">50K", .n = 20))

Save metadata to a JSON file

Description

Save metadata to a JSON file

Usage

save_metadata(meta, path)

Arguments

Value

Invisibly returns meta.

Examples

meta <- metadata() |> set_column_type("age", "numerical")
tmp <- tempfile(fileext = ".json")
save_metadata(meta, tmp)
meta2 <- load_metadata(tmp)

Set the type of a column in metadata

Description

Set the type of a column in metadata

Usage

set_column_type(meta, column, type)

Arguments

Value

The updated rsdv_metadata object (for piping).

Examples

metadata() |> set_column_type("age", "numerical")

Set the primary key column of the metadata

Description

Set the primary key column of the metadata

Usage

set_primary_key(meta, column)

Arguments

Value

The updated rsdv_metadata object (for piping).

Examples

meta <- metadata() |>
  set_column_type("id", "id") |>
  set_primary_key("id")

Total variation distance similarity score per categorical column

Description

Total variation distance similarity score per categorical column

Usage

tvd_similarity(real, synthetic, meta)

Arguments

Value

A tibble with columns column (chr) and score (dbl, 0–1, higher = better).

Examples

syn   <- gaussian_copula_synthesizer(metadata(adult_income)) |> fit(adult_income)
synth <- sample(syn, n = 500)
tvd_similarity(adult_income, synth, metadata(adult_income))

Validate that a data frame is compatible with metadata

Description

Checks that all columns registered in meta are present in data.

Usage

validate_data(data, meta)

Arguments

Value

Invisibly TRUE; throws an error if validation fails.

Examples

meta <- metadata() |> set_column_type("age", "numerical")
validate_data(data.frame(age = 1:5), meta)