Compare original and synthetic datasets

Compares an original dataset with its synthetic double across dataset-level dimensions, numeric distributions, categorical distributions, and numeric correlations. Returns a structured dataganger_comparison object.

Usage

compare_synthetic(original, synthetic, roles = NULL)

Arguments

original

The original data frame.

synthetic

The synthetic data frame (from synthesize_data()).

roles

Optional; a dataganger_roles object from detect_roles().

Value

An S3 object of class dataganger_comparison, a list with components dataset, numeric, categorical, relationship, interaction, utility (a global pMSE-based utility diagnostic; see Details), disclosure (role-aware disclosure risk diagnostics; see Details), privacy_flags, and meta.

Details

utility reports the propensity-score pMSE / S_pMSE utility diagnostic of Snoke, Raab, Nowok, Dibben & Slavkovic (2018), the same formula synthpop::utility.gen() uses for its logistic-regression method: S_pMSE near 1 means a model fit to distinguish original from synthetic rows does no better than chance on the shared predictor columns (high utility for that joint distribution); higher values mean the two datasets are more detectably different. This is a utility measure, not a privacy measure – it says nothing about disclosure risk.

disclosure evaluates role-aware disclosure risk diagnostics using synthpop. Identity disclosure is assessed via replicated uniques (synthpop::replicated.uniques()) across quasi-identifier keys identified in roles: original uniques replicated in synthetic data indicate potential re-identification risk. When sensitive target attributes are present, attribute disclosure risk is evaluated via the DiSCO measure (synthpop::disclosure()).

Examples

dat <- data.frame(x = 1:10, y = letters[1:10])
spec <- synth_spec(purpose = "demo")
syn <- synthesize_data(dat, spec)
#> Warning: Cannot guarantee level presence for columns: y (10 levels at k = 5 require
#> minimum n = 50). Largest minimum n = 50; output has n = 10. Restoring as many
#> levels as fit without removing another level's last copy.
compare_synthetic(dat, syn)
#> ℹ Not enough numeric columns (1) for correlation comparison.
#>   Need at least 2 numeric columns with non-zero variance.
#> 
#> ── DataGangeR Comparison ───────────────────────────────────────────────────────
#> 
#> ── Dataset ──
#> 
#> • Rows: 10 (original) -> 10 (synthetic)
#> • Columns: 2 (original) -> 2 (synthetic)
#> • Type match: 50%
#> • Missing: 0% (original) -> 0% (synthetic)
#> 
#> ── Numeric -- top 3 by |standardized difference| ──
#> 
#> • x: std diff = -0.429
#> Orig mean (SD): 5.5 (3.03)
#> 
#> ── Categorical -- top 3 by distributional difference ──
#> 
#> • y: p = 0.395, TVD = 1
#> Levels: 10 (orig) -> 10 (syn)
#> 
#> ── Utility ──
#> 
#> • S_pMSE = 2.11 (1.0 = the model could not tell original and synthetic rows
#> apart on 19 predictors; higher = more detectably different)
#> This is a utility measure, not a privacy measure -- a low score here is not
#> evidence the data are safe to release.
#> 
#> ── Disclosure risk (synthpop) ──
#> 
#> • Not computed: Roles must be provided to determine quasi-identifier keys and
#> sensitive targets.