Synthetic Data Generation Workflow - Bank Marketing Data

---

Synthetic data is artificially generated data that mimics the statistical properties of a real dataset (e.g. column structure, distributions, relationships between variables). It is useful for sharing data safely, augmenting datasets, and testing pipelines without exposing sensitive information.

This notebook shows a synthetic data generation workflow using the Synthetic Data Vault (SDV) library and a marketing dataset.

The dataset represents a bank marketing campaign from UCI Machine Learning Repository, with 4,119 records, 21 variables, and mixed numeric and categorical types.

The goal is to generate a synthetic version of the dataset that preserves the statistical properties of the original, including column distributions, inter-variable correlations, and category adherence.

The final model used in this workflow is SDV's CopulaGANSynthesizer, a hybrid approach that combines Gaussian copula marginal fitting with CTGAN-based joint distribution learning. Four synthesizers were evaluated in total; see the Synthetic Data Generation section for a discussion of the tradeoffs.

The notebook is organized as follows:

• Data Description: discussion of context, interesting variables, and missing values
• Data Loading: schema definition, ingestion, and data summary
• Synthetic Data Generation: metadata detection, model training, and sampling
• Synthetic Data Validation: distribution comparison, correlation analysis, quality scores, and constraint checks
• Save Outputs: save synthetic dataset, fitted model, and metadata to disk
• Conditional Generation: brief demo of SDV's conditional generation functionality

Data Description

---

The dataset comes from direct marketing campaigns run by a Portuguese bank between 2008 and 2013. Each record represents one client contact (by phone), with the target variable y indicating whether the client subscribed to a term deposit.

Variables fall into three groups:

• Client attributes: age, job, marital status, education, and whether the client has credit in default, a housing loan, or a personal loan
• Campaign contact info: contact type, month, day of week, call duration, and number of contacts in the current campaign
• Economic context: employment variation rate, consumer price index, consumer confidence index, Euribor 3-month rate, and number of employees

Note on pdays: This variable records the number of days since the client was last contacted in a previous campaign. The value 999 means they were never previously contacted, which accounts for about 96% of rows. This is a sentinel value, not missing data, and it makes pdays effectively bimodal, which will show up in the validation results.

Note on missing values: Some columns (e.g. default, housing, and loan) contain an unknown category which we treat as a valid category rather than imputing or converting to NaN. This preserves the information that the value was unknown at the time of contact. Otherwise, the dataset has no true missing values.

Data Loading

---

We define a schema with data types for each column and then load the data.

import pandas as pd
import numpy as np
from minieda import summarize
import sdv
from sdv.metadata import Metadata
from sdv.single_table import CopulaGANSynthesizer
from sdv.evaluation.single_table import evaluate_quality
from sdv.evaluation.single_table import run_diagnostic

from sdmetrics.single_table import NewRowSynthesis
from sdmetrics.single_table import DCRBaselineProtection

from sdv.sampling import Condition

import matplotlib.pyplot as plt
import os
import warnings

warnings.filterwarnings('ignore', category=FutureWarning)
warnings.filterwarnings('ignore', category=UserWarning)

print(pd.__version__)
print(sdv.__version__)

2.2.3
1.34.2

# Function for printing tables
def style_table(df, caption, precision=1, caption_side="bottom", font_size="16px", font_weight="normal"):
    return (df.style
            .format(precision=precision).set_caption(caption)
            .set_table_styles([{"selector": "caption", 
                                "props": [("caption-side", caption_side),
                                          ("font-size", font_size),
                                          ("font-weight", font_weight)]}]))

data_schema = {
    'age':              'int64',      # client age
    'job':              'category',   # type of job (admin, blue-collar, entrepreneur, housemaid, management, retired, self-employed, services, student, technician, unemployed, unknown)
    'marital':          'category',   # marital status (divorced, married, single, unknown)
    'education':        'category',   # education level (basic.4y, basic.6y, basic.9y, high.school, illiterate, professional.course, university.degree, unknown)
    'default':          'category',   # has credit in default? (yes, no, unknown)
    'housing':          'category',   # has housing loan? (yes, no, unknown)
    'loan':             'category',   # has personal loan? (yes, no, unknown)
    'contact':          'category',   # contact communication type (cellular, telephone)
    'month':            'category',   # last contact month of year (jan–dec)
    'day_of_week':      'category',   # last contact day of the week (mon, tue, wed, thu, fri)
    'duration':         'int64',      # last contact duration in seconds — benchmark only, not for predictive modeling
    'campaign':         'int64',      # number of contacts performed during this campaign (includes last contact)
    'pdays':            'int64',      # days since client was last contacted from previous campaign (999 = not previously contacted)
    'previous':         'int64',      # number of contacts performed before this campaign
    'poutcome':         'category',   # outcome of previous marketing campaign (failure, nonexistent, success)
    'emp.var.rate':     'float64',    # employment variation rate — quarterly indicator
    'cons.price.idx':   'float64',    # consumer price index — monthly indicator
    'cons.conf.idx':    'float64',    # consumer confidence index — monthly indicator
    'euribor3m':        'float64',    # euribor 3 month rate — daily indicator
    'nr.employed':      'float64',    # number of employees — quarterly indicator
    'y':                'category',   # target: has the client subscribed a term deposit? (yes, no)
}

# Load data with schema
data = pd.read_csv('data/bank-additional.csv', sep=";", dtype=data_schema)

# Convert yes/no columns to True/False
bool_cols = ['y']
for col in bool_cols:
    data[col] = data[col].map({'yes': True, 'no': False})

summarize(data)

	dtype	count	unique	unique_pct	missing	missing_pct	zero	zero_pct	top	freq	mean	std	min	50%	max	skew
age	int64	4119	67	1.63	0	0.0	0	0.00			40.11	10.31	18.0	38.0	88.0	0.72
campaign	int64	4119	25	0.61	0	0.0	0	0.00			2.54	2.57	1.0	2.0	35.0	4.0
nr.employed	float64	4119	11	0.27	0	0.0	0	0.00			5166.48	73.67	4963.6	5191.0	5228.1	-1.08
euribor3m	float64	4119	234	5.68	0	0.0	0	0.00			3.62	1.73	0.64	4.86	5.04	-0.72
cons.conf.idx	float64	4119	26	0.63	0	0.0	0	0.00			-40.5	4.59	-50.8	-41.8	-26.9	0.29
cons.price.idx	float64	4119	26	0.63	0	0.0	0	0.00			93.58	0.58	92.2	93.75	94.77	-0.22
emp.var.rate	float64	4119	10	0.24	0	0.0	0	0.00			0.08	1.56	-3.4	1.1	1.4	-0.73
previous	int64	4119	7	0.17	0	0.0	3523	85.53			0.19	0.54	0.0	0.0	6.0	4.02
pdays	int64	4119	21	0.51	0	0.0	2	0.05			960.42	191.92	0.0	999.0	999.0	-4.78
duration	int64	4119	828	20.10	0	0.0	1	0.02			256.79	254.7	0.0	181.0	3643.0	3.29
job	category	4119	12	0.29	0	0.0	0	0.00	admin.	1012
day_of_week	category	4119	5	0.12	0	0.0	0	0.00	thu	860
month	category	4119	10	0.24	0	0.0	0	0.00	may	1378
contact	category	4119	2	0.05	0	0.0	0	0.00	cellular	2652
poutcome	category	4119	3	0.07	0	0.0	0	0.00	nonexistent	3523
loan	category	4119	3	0.07	0	0.0	0	0.00	no	3349
housing	category	4119	3	0.07	0	0.0	0	0.00	yes	2175
default	category	4119	3	0.07	0	0.0	0	0.00	no	3315
education	category	4119	8	0.19	0	0.0	0	0.00	university.degree	1264
marital	category	4119	4	0.10	0	0.0	0	0.00	married	2509
y	category	4119	2	0.05	0	0.0	0	0.00	False	3668

Synthetic Data Generation

---

We use CopulaGANSynthesizer as the primary synthesizer. It combines Gaussian copula marginal fitting with CTGAN-based joint distribution learning, which produces better fidelity on both column distributions and categorical relationships than GaussianCopulaSynthesizer alone (Column Shapes 0.88 vs 0.74, categorical pair trends 0.67 vs 0.43).

Three other synthesizers were evaluated (code and results not shown in this notebook):

• GaussianCopulaSynthesizer produced good validity and privacy scores but systematically miscalibrated heavily imbalanced categorical columns, inflating the unknown category in default and marital to become the dominant value in the synthetic data.
• CTGANSynthesizer failed to reproduce inter-variable correlations at this dataset size, producing a near-flat correlation matrix.
• TVAESynthesizer dropped minority categories entirely for several columns including loan and marital.

CopulaGAN avoids all three failure modes while accepting a noticeable tradeoff in nearest-neighbor privacy score (0.37 vs 0.53 for Gaussian Copula) — synthetic rows sit closer to real rows than random noise.

We sample 4,119 rows (same as the original dataset).

metadata = Metadata.detect_from_dataframe(data=data, table_name='bank_marketing')

# CopulaGAN requires categorical columns to be `object` dtype instead of pandas' `category`
data_copulagan = data.copy()
cat_cols = data.select_dtypes('category').columns
data_copulagan[cat_cols] = data_copulagan[cat_cols].astype(object)

# Instantiate and fit the model
synthesizer = CopulaGANSynthesizer(metadata, epochs=150, verbose=True)
synthesizer.fit(data_copulagan)

# Generate synthetic data, with the same number of rows as the original dataset
synthetic_data = synthesizer.sample(num_rows=data_copulagan.shape[0])

# Restore category dtype to match real data (CopulaGAN outputs object dtype)
synthetic_data[cat_cols] = synthetic_data[cat_cols].astype('category')

print(synthetic_data.shape)

Gen. (-00.64) | Discrim. (-00.49): 100%|█████████████████████████████████████████████| 150/150 [01:39<00:00,  1.51it/s]

(4119, 21)

synthetic_data.head()

	age	job	marital	education	default	housing	loan	contact	month	day_of_week	...	campaign	pdays	previous	poutcome	emp.var.rate	cons.price.idx	cons.conf.idx	euribor3m	nr.employed	y
0	49	technician	single	university.degree	unknown	no	no	cellular	jul	mon	...	7	999	0	nonexistent	1.4	94.500	-42.3	4.967	5228.1	False
1	50	blue-collar	married	basic.4y	no	yes	no	telephone	jun	mon	...	2	999	0	nonexistent	1.4	93.950	-36.3	4.145	5228.1	False
2	50	blue-collar	single	basic.6y	no	yes	no	cellular	jul	tue	...	6	999	0	nonexistent	1.2	93.546	-35.9	4.955	5228.1	False
3	36	technician	married	basic.9y	no	yes	yes	cellular	jul	wed	...	4	999	0	nonexistent	1.4	94.097	-36.7	4.980	5228.1	False
4	88	admin.	single	university.degree	no	yes	no	cellular	apr	thu	...	1	4	1	failure	-1.7	93.007	-26.9	1.085	5007.6	True

5 rows × 21 columns

summarize(synthetic_data)

	dtype	count	unique	unique_pct	missing	missing_pct	zero	zero_pct	top	freq	mean	std	min	50%	max	skew
age	int64	4119	68	1.65	0	0.0	0	0.00			47.81	14.07	20.0	46.0	88.0	0.73
campaign	int64	4119	33	0.80	0	0.0	0	0.00			3.16	4.15	1.0	2.0	34.0	3.75
nr.employed	float64	4119	1453	35.28	0	0.0	0	0.00			5142.09	86.9	4963.6	5178.9	5228.1	-0.72
euribor3m	float64	4119	1462	35.49	0	0.0	0	0.00			3.36	1.73	0.64	4.31	4.99	-0.36
cons.conf.idx	float64	4119	221	5.37	0	0.0	0	0.00			-39.34	5.6	-50.8	-41.3	-26.9	0.56
cons.price.idx	float64	4119	1392	33.79	0	0.0	0	0.00			93.56	0.59	92.2	93.49	94.77	-0.03
emp.var.rate	float64	4119	49	1.19	0	0.0	19	0.46			-0.15	1.65	-3.4	0.7	1.4	-0.55
previous	int64	4119	7	0.17	0	0.0	3090	75.02			0.41	0.92	0.0	0.0	6.0	3.2
pdays	int64	4119	190	4.61	0	0.0	59	1.43			858.2	336.89	0.0	999.0	999.0	-1.99
duration	int64	4119	1240	30.10	0	0.0	0	0.00			482.64	546.12	4.0	307.0	3643.0	2.83
job	category	4119	12	0.29	0	0.0	0	0.00	blue-collar	889
day_of_week	category	4119	5	0.12	0	0.0	0	0.00	tue	1166
month	category	4119	10	0.24	0	0.0	0	0.00	may	920
contact	category	4119	2	0.05	0	0.0	0	0.00	cellular	3016
poutcome	category	4119	3	0.07	0	0.0	0	0.00	nonexistent	3178
loan	category	4119	3	0.07	0	0.0	0	0.00	no	2695
housing	category	4119	3	0.07	0	0.0	0	0.00	yes	2561
default	category	4119	3	0.07	0	0.0	0	0.00	no	3186
education	category	4119	8	0.19	0	0.0	0	0.00	high.school	1153
marital	category	4119	4	0.10	0	0.0	0	0.00	married	1927
y	category	4119	2	0.05	0	0.0	0	0.00	False	3168

Synthetic Data Validation

---

In this section, we use several methods to check how the synthetic data compares to the original dataset.

Per-variable Distribution Comparison

We plot each column of the real and synthetic data side-by-side using KDE curves for numeric columns and grouped bar plots for categorical columns.

A well-fitting synthetic dataset should show close overlap throughout.

The table below the plots summarizes mean, std, min, and max for the numeric columns.

There is a noticeable difference in the means of several features, e.g. duration, previous, emp.var.rate.

categorical_cols = ['job', 'marital', 'education', 'default', 'housing', 'loan',
                    'contact', 'month', 'day_of_week', 'poutcome']
boolean_cols     = ['y']
numeric_cols     = ['age', 'duration', 'campaign', 'pdays', 'previous',
                    'emp.var.rate', 'cons.price.idx', 'cons.conf.idx',
                    'euribor3m', 'nr.employed']

all_cols = categorical_cols + boolean_cols + numeric_cols
n_cols   = len(all_cols)

# Set up grid
n_grid_cols = 3
n_grid_rows = int(np.ceil(n_cols / n_grid_cols))

fig, axes = plt.subplots(n_grid_rows, n_grid_cols, figsize=(18, n_grid_rows * 4))
axes = axes.flatten()

for i, col in enumerate(all_cols):
    ax = axes[i]

    if col in numeric_cols:
        # Overlapping KDE plots for numeric columns
        data[col].plot.kde(ax=ax, label='Real', color='steelblue', linewidth=2)
        synthetic_data[col].plot.kde(ax=ax, label='Synthetic', color='coral',
                                      linewidth=2, linestyle='--')
        ax.set_title(col, fontsize=13)
        ax.legend()
        ax.tick_params(axis='both', labelsize=11)

    else:
        # Grouped bar chart for categorical and boolean columns
        real_counts = data[col].value_counts(normalize=True).sort_index()
        synth_counts = synthetic_data[col].value_counts(normalize=True).sort_index()

        # Align categories in case synthetic has different order
        all_categories = sorted(set(real_counts.index) | set(synth_counts.index),
                                key=str)
        real_counts  = real_counts.reindex(all_categories, fill_value=0)
        synth_counts = synth_counts.reindex(all_categories, fill_value=0)

        x = np.arange(len(all_categories))
        width = 0.35

        ax.bar(x - width/2, real_counts.values,  width, label='Real',
               color='steelblue', alpha=0.8)
        ax.bar(x + width/2, synth_counts.values, width, label='Synthetic',
               color='coral', alpha=0.8)

        ax.set_xticks(x)
        ax.set_xticklabels(all_categories, rotation=45, ha='right', fontsize=11)
        ax.set_title(col, fontsize=13)
        ax.legend()

# Hide any unused subplots
for j in range(i + 1, len(axes)):
    axes[j].set_visible(False)

plt.suptitle('Per-Variable Distribution: Real vs Synthetic', fontsize=16, y=1.02)
plt.tight_layout()
plt.show()

real_stats  = data[numeric_cols].describe().T[['mean', 'std', 'min', 'max']]
synth_stats = synthetic_data[numeric_cols].describe().T[['mean', 'std', 'min', 'max']]

real_stats.columns  = ['real_mean', 'real_std', 'real_min', 'real_max']
synth_stats.columns = ['synth_mean', 'synth_std', 'synth_min', 'synth_max']

summary = pd.concat([real_stats, synth_stats], axis=1)
style_table(summary.round(3), caption="Numeric Column Statistics: Real vs Synthetic")

Numeric Column Statistics: Real vs Synthetic

	real_mean	real_std	real_min	real_max	synth_mean	synth_std	synth_min	synth_max
age	40.1	10.3	18.0	88.0	47.8	14.1	20.0	88.0
duration	256.8	254.7	0.0	3643.0	482.6	546.1	4.0	3643.0
campaign	2.5	2.6	1.0	35.0	3.2	4.1	1.0	34.0
pdays	960.4	191.9	0.0	999.0	858.2	336.9	0.0	999.0
previous	0.2	0.5	0.0	6.0	0.4	0.9	0.0	6.0
emp.var.rate	0.1	1.6	-3.4	1.4	-0.1	1.6	-3.4	1.4
cons.price.idx	93.6	0.6	92.2	94.8	93.6	0.6	92.2	94.8
cons.conf.idx	-40.5	4.6	-50.8	-26.9	-39.3	5.6	-50.8	-26.9
euribor3m	3.6	1.7	0.6	5.0	3.4	1.7	0.6	5.0
nr.employed	5166.5	73.7	4963.6	5228.1	5142.1	86.9	4963.6	5228.1

Correlation Structure

---

Beyond individual column distributions, a synthetic dataset should preserve relationships between variables. We first compare numeric correlation matrices visually using heatmaps. Then we use SDV's evaluate_quality function to compute Column Pair Trends scores.

Column Pair Trends evaluates two types of relationships:

• Numeric-numeric pairs (Correlation Similarity): how well linear correlations between numeric columns are reproduced
• Categorical/mixed pairs (Contingency Similarity): how well associations between categorical columns, and between categorical and numeric columns, are preserved

Each row in the tables below represents a column pair. The Score measures how well the relationship was preserved in the synthetic data, from 0 (not preserved) to 1 (perfectly preserved). The tables below show only column pairs that meet SDV's association threshold; pairs where the real data shows a weak relationship are excluded.

We also compute mean scores to summarize performance of each of the pair types.

fig, axes = plt.subplots(1, 2, figsize=(12, 6))

real_corr  = data[numeric_cols].corr()
synth_corr = synthetic_data[numeric_cols].corr()

# Shared color scale
vmin, vmax = -1, 1

for ax, corr, title in zip(axes, [real_corr, synth_corr], ['Real Data', 'Synthetic Data']):
    im = ax.imshow(corr, cmap='coolwarm', vmin=vmin, vmax=vmax, aspect='auto')
    ax.set_xticks(range(len(numeric_cols)))
    ax.set_yticks(range(len(numeric_cols)))
    ax.set_xticklabels(numeric_cols, rotation=45, ha='right', fontsize=11)
    ax.set_yticklabels(numeric_cols, fontsize=11)
    ax.set_title(title, fontsize=14)

    # Annotate cells with correlation values
    for i in range(len(numeric_cols)):
        for j in range(len(numeric_cols)):
            ax.text(j, i, f'{corr.iloc[i, j]:.2f}',
                    ha='center', va='center', fontsize=8,
                    color='white' if abs(corr.iloc[i, j]) > 0.6 else 'black')

# Dedicated colorbar axis
cbar_ax = fig.add_axes([0.92, 0.18, 0.02, 0.7])  # [left, bottom, width, height]
fig.colorbar(im, cax=cbar_ax)

plt.suptitle('Correlation Matrix: Real vs Synthetic', fontsize=16)
plt.tight_layout(rect=[0, 0, 0.91, 1])
plt.show()

quality_report = evaluate_quality(
    real_data=data,
    synthetic_data=synthetic_data,
    metadata=metadata,
    verbose=False
)

pair_trends = quality_report.get_details(property_name='Column Pair Trends')
pair_trends_nona = pair_trends.dropna(subset=['Score'])

# Split into correlation (numeric-numeric) and contingency (categorical) pairs
correlation_pairs = pair_trends_nona[pair_trends_nona['Metric'] == 'CorrelationSimilarity']
contingency_pairs = pair_trends_nona[pair_trends_nona['Metric'] == 'ContingencySimilarity']

style_table(correlation_pairs[['Column 1', 'Column 2', 'Score', 'Real Correlation', 'Synthetic Correlation']],
            caption="Numeric-Numeric Pairs (Correlation Similarity)")

Numeric-Numeric Pairs (Correlation Similarity)

	Column 1	Column 2	Score	Real Correlation	Synthetic Correlation
195	emp.var.rate	cons.price.idx	0.8	0.8	0.4
197	emp.var.rate	euribor3m	0.8	1.0	0.5
198	emp.var.rate	nr.employed	0.8	0.9	0.4
201	cons.price.idx	euribor3m	0.9	0.7	0.4
207	euribor3m	nr.employed	0.8	0.9	0.5

style_table(contingency_pairs[['Column 1', 'Column 2', 'Score', 'Real Association']],
            caption="Categorical/Mixed Pairs (Contingency Similarity)")

Categorical/Mixed Pairs (Contingency Similarity)

	Column 1	Column 2	Score	Real Association
21	job	education	0.6	0.4
90	housing	loan	0.8	0.7
119	contact	month	0.7	0.6
126	contact	emp.var.rate	0.6	0.7
127	contact	cons.price.idx	0.6	0.7
128	contact	cons.conf.idx	0.7	0.4
129	contact	euribor3m	0.8	0.5
130	contact	nr.employed	0.7	0.5
138	month	emp.var.rate	0.4	0.7
139	month	cons.price.idx	0.2	0.7
140	month	cons.conf.idx	0.3	0.6
141	month	euribor3m	0.5	0.5
142	month	nr.employed	0.4	0.6
164	duration	y	0.7	0.4
174	pdays	previous	0.8	0.6
175	pdays	poutcome	0.8	0.9
181	pdays	y	0.8	0.3
182	previous	poutcome	0.8	0.7
189	poutcome	emp.var.rate	0.6	0.4
190	poutcome	cons.price.idx	0.6	0.4
191	poutcome	cons.conf.idx	0.7	0.4
192	poutcome	euribor3m	0.7	0.4
193	poutcome	nr.employed	0.6	0.4
194	poutcome	y	0.9	0.3
199	emp.var.rate	y	0.7	0.3
203	cons.price.idx	y	0.6	0.3
206	cons.conf.idx	y	0.7	0.4
208	euribor3m	y	0.7	0.4
209	nr.employed	y	0.7	0.4

print(f"Numeric-Numeric Pairs Mean Score: {correlation_pairs['Score'].mean():.3f}")
print(f"Categorical/Mixed Pairs Mean Score: {contingency_pairs['Score'].mean():.3f}")

Numeric-Numeric Pairs Mean Score: 0.798
Categorical/Mixed Pairs Mean Score: 0.644

SDV Quality Scores

---

SDV's evaluate_quality function scores each column individually on how well its distribution was reproduced ('Column Shapes') and each column pair on how well their relationship was preserved ('Column Pair Trends').

We already examined the pair trends above; here we look at the per-column scores.

quality_report.get_details(property_name='Column Shapes')

	Column	Metric	Score
0	age	KSComplement	0.764020
1	job	TVComplement	0.879825
2	marital	TVComplement	0.858704
3	education	TVComplement	0.866472
4	default	TVComplement	0.968682
5	housing	TVComplement	0.873513
6	loan	TVComplement	0.841224
7	contact	TVComplement	0.911629
8	month	TVComplement	0.871328
9	day_of_week	TVComplement	0.837096
10	duration	KSComplement	0.749697
11	campaign	KSComplement	0.932993
12	pdays	KSComplement	0.889051
13	previous	TVComplement	0.894877
14	poutcome	TVComplement	0.916242
15	emp.var.rate	KSComplement	0.864530
16	cons.price.idx	KSComplement	0.817674
17	cons.conf.idx	KSComplement	0.815246
18	euribor3m	KSComplement	0.837825
19	nr.employed	KSComplement	0.738043
20	y	TVComplement	0.878611

Privacy and Memorization Checks

---

We run two checks to confirm that the synthetic data doesn't reproduce records from the training data.

• New Row Score (NewRowSynthesis): checks whether any synthetic rows are exact copies of real rows. A score of 1.0 means every synthetic row is new.
• Nearest-Neighbor Privacy Score (DCRBaselineProtection): measures how close synthetic rows are to their nearest real neighbors, relative to a random data baseline. A score of 1.0 means the synthetic data is no closer to the real data than random noise (the best possible privacy). A score near 0.5 is typical for a well-trained model.

The New Row Score of 1.0 confirms no exact copies. The Nearest-Neighbor Privacy Score of 0.354 is below the ideal 0.5 baseline — the synthetic data has a median distance of 0.126 to its nearest real neighbors, compared to 0.356 for randomly generated data. In other words, synthetic rows are closer to real rows than random noise would be, which is expected for a model that has learned the data distribution well. This does not indicate memorization of individual records, which is confirmed by the New Row Score of 1.0.

metadata_dict = metadata.to_dict()['tables']['bank_marketing']

def to_str_categories(df):
    return df.copy().assign(**{
        col: df[col].astype(str) 
        for col in df.select_dtypes('category').columns
    })

new_row_score = NewRowSynthesis.compute(
    real_data=to_str_categories(data),
    synthetic_data=to_str_categories(synthetic_data),
    metadata=metadata_dict,
    numerical_match_tolerance=0.01,
    synthetic_sample_size=None
)

nn_privacy_score = DCRBaselineProtection.compute_breakdown(
    real_data=data,
    synthetic_data=synthetic_data,
    metadata=metadata_dict
)

print(f"New Row Score: {new_row_score:.3f}")
print(f"Nearest-Neighbor Privacy Score: {nn_privacy_score['score']:.3f}")
print(f"Synthetic median DCR to real data: {nn_privacy_score['median_DCR_to_real_data']['synthetic_data']:.3f}")
print(f"Random baseline median DCR to real data: {nn_privacy_score['median_DCR_to_real_data']['random_data_baseline']:.3f}")

New Row Score: 1.000
Nearest-Neighbor Privacy Score: 0.417
Synthetic median DCR to real data: 0.148
Random baseline median DCR to real data: 0.354

SDV Validity Checks

---

To conclude this section, we run SDV's built-in diagnostic report, which captures basic diagnostic measurements across the entire dataset, reporting areas that may be problematic. It checks:

• Data Structure: the synthetic table has the same column names as the real data.
• Data Validity: synthetic values respect the original data's column types and value ranges.

We expect both the scores to be close to 100%.

diagnostic_report = run_diagnostic(
    real_data=data,
    synthetic_data=synthetic_data,
    metadata=metadata,
    verbose=True
)

Generating report ...

(1/2) Evaluating Data Validity: |███████████████████████████████████████████████████| 21/21 [00:00<00:00, 1613.70it/s]|
Data Validity Score: 100.0%

(2/2) Evaluating Data Structure: |█████████████████████████████████████████████████████| 1/1 [00:00<00:00, 333.41it/s]|
Data Structure Score: 100.0%

Overall Score (Average): 100.0%

Synthetic Data Quality Summary

---

The table below summarizes the key validation metrics computed in this section.

• Data Validity and Data Structure should score close to 100%.
• Column Shapes, Column Pair Trends, and the privacy metrics don't have a fixed target; however, higher is better for quality scores, while a Nearest-Neighbor Privacy Score near 0.5 indicates a good balance between utility and privacy.

summary = {
    'Column Shapes':                    quality_report.get_properties().set_index('Property').loc['Column Shapes', 'Score'],
    'Column Pair Trends (numeric)':     correlation_pairs['Score'].mean(),
    'Column Pair Trends (categorical)': contingency_pairs['Score'].mean(),
    'Data Validity':                    diagnostic_report.get_properties().set_index('Property').loc['Data Validity', 'Score'],
    'Data Structure':                   diagnostic_report.get_properties().set_index('Property').loc['Data Structure', 'Score'],
    'New Row Score':                    new_row_score,
    'Nearest-Neighbor Privacy':         nn_privacy_score['score'],
}

summary_df = pd.DataFrame.from_dict(summary, orient='index', columns=['Score'])
style_table(summary_df, caption='Synthetic Data Quality Summary', precision=2)

Synthetic Data Quality Summary

	Score
Column Shapes	0.86
Column Pair Trends (numeric)	0.80
Column Pair Trends (categorical)	0.64
Data Validity	1.00
Data Structure	1.00
New Row Score	1.00
Nearest-Neighbor Privacy	0.42

Save Outputs

---

For reproducibility, we save the following items:

• The synthetic dataset
• Fitted synthesizer
• Metadata

The synthesizer can be reloaded to generate additional synthetic rows without retraining.

os.makedirs('outputs', exist_ok=True)

# Save synthetic data to CSV
synthetic_data.to_csv('outputs/synthetic_data.csv', index=False)

# Save the fitted synthesizer for reproducibility
synthesizer.save('outputs/copulagan_synthesizer.pkl')

# Save metadata to JSON
metadata.save_to_json('outputs/metadata.json', mode='overwrite')

Conditional Generation

---

Beyond generating a full synthetic dataset, SDV supports conditional generation: we sample rows that match specified demographic or behavioral criteria while drawing the remaining columns from the learned distribution. This can be useful for generating targeted test datasets, augmenting underrepresented segments, or simulating specific customer profiles.

Conditional generation respects statistical patterns learned from the data but does not enforce logical constraints between columns. In practice the model learns these relationships reasonably well. For example, conditioning on job='retired' produces a mean age of 39 with plausible older clients represented. However, implausible combinations can still appear, such as retired clients in their late 20s. SDV's constraint framework can address this for cases where hard rules are required, but is outside the scope of this notebook.

# Basic example: generate 200 rows for young clients (age 25)
young_condition = Condition(
    num_rows=200,
    column_values={'age': 25}
)
young_synthetic = synthesizer.sample_from_conditions(conditions=[young_condition])
young_synthetic[cat_cols] = young_synthetic[cat_cols].astype('category') # Restore category dtype to match real data (CopulaGAN outputs object dtype)
print(young_synthetic.shape)
young_synthetic.head()

Sampling conditions: 100%|███████████████████████████████████████████████████████████| 200/200 [00:04<00:00, 48.81it/s]

(200, 21)

	age	job	marital	education	default	housing	loan	contact	month	day_of_week	...	campaign	pdays	previous	poutcome	emp.var.rate	cons.price.idx	cons.conf.idx	euribor3m	nr.employed	y
0	25	blue-collar	married	university.degree	unknown	yes	unknown	telephone	may	thu	...	2	999	0	failure	1.0	93.997	-36.5	4.928	5169.5	False
1	25	unemployed	single	university.degree	no	no	yes	cellular	mar	wed	...	2	999	0	nonexistent	-1.1	93.548	-50.8	1.560	5091.6	True
2	25	blue-collar	married	unknown	unknown	no	yes	cellular	jul	tue	...	2	999	0	nonexistent	1.4	94.004	-42.9	4.962	5228.1	False
3	25	admin.	single	basic.9y	unknown	unknown	no	cellular	sep	fri	...	6	999	0	nonexistent	-1.7	93.955	-42.0	4.846	5228.1	False
4	25	blue-collar	divorced	high.school	no	yes	no	cellular	may	fri	...	2	999	0	nonexistent	1.4	94.455	-42.0	4.972	5228.1	False

5 rows × 21 columns

# Complex example: generate 200 rows for retired clients, contacted by cellular in May, with no previous campaign contact
complex_condition = Condition(
    num_rows=200,
    column_values={
        'job':      'retired',
        'contact':  'cellular',
        'month':    'may',
        'poutcome': 'nonexistent',
        'previous': 0
    }
)
complex_synthetic = synthesizer.sample_from_conditions(conditions=[complex_condition])
complex_synthetic[cat_cols] = complex_synthetic[cat_cols].astype('category') # Restore category dtype to match real data (CopulaGAN outputs object dtype)
print(complex_synthetic.shape)
complex_synthetic.head()

Sampling conditions: 100%|███████████████████████████████████████████████████████████| 200/200 [00:07<00:00, 28.25it/s]

(200, 21)

	age	job	marital	education	default	housing	loan	contact	month	day_of_week	...	campaign	pdays	previous	poutcome	emp.var.rate	cons.price.idx	cons.conf.idx	euribor3m	nr.employed	y
0	33	retired	married	basic.9y	no	no	no	cellular	may	tue	...	1	999	0	nonexistent	0.6	94.559	-35.9	4.453	5228.1	False
1	50	retired	married	basic.9y	unknown	yes	no	cellular	may	fri	...	8	999	0	nonexistent	0.1	94.122	-41.8	4.915	5194.5	False
2	61	retired	married	professional.course	no	yes	yes	cellular	may	tue	...	1	999	0	nonexistent	1.4	93.980	-47.2	1.141	4989.6	True
3	44	retired	married	university.degree	no	no	yes	cellular	may	wed	...	2	999	0	nonexistent	1.4	93.095	-42.7	4.966	5228.1	False
4	45	retired	married	unknown	unknown	yes	no	cellular	may	fri	...	1	999	0	nonexistent	1.4	94.059	-42.9	4.890	5191.3	False

5 rows × 21 columns

summarize(complex_synthetic)

	dtype	count	unique	unique_pct	missing	missing_pct	zero	zero_pct	top	freq	mean	std	min	50%	max	skew
age	int64	200	47	23.5	0	0.0	0	0.0			46.22	12.19	24.0	45.0	88.0	0.92
campaign	int64	200	15	7.5	0	0.0	0	0.0			2.91	3.96	1.0	2.0	30.0	4.49
nr.employed	float64	200	139	69.5	0	0.0	0	0.0			5138.65	87.1	4963.6	5177.45	5228.1	-0.79
euribor3m	float64	200	168	84.0	0	0.0	0	0.0			3.46	1.67	0.64	4.32	4.99	-0.53
cons.conf.idx	float64	200	80	40.0	0	0.0	0	0.0			-40.49	4.91	-50.8	-41.75	-26.9	0.74
cons.price.idx	float64	200	182	91.0	0	0.0	0	0.0			93.51	0.59	92.2	93.46	94.77	-0.08
emp.var.rate	float64	200	46	23.0	0	0.0	3	1.5			-0.45	1.73	-3.4	-0.4	1.4	-0.37
previous	int64	200	1	0.5	0	0.0	200	100.0			0.0	0.0	0.0	0.0	0.0	0.0
pdays	int64	200	6	3.0	0	0.0	1	0.5			974.51	153.37	0.0	999.0	999.0	-6.14
duration	int64	200	183	91.5	0	0.0	0	0.0			492.77	531.43	5.0	327.0	3643.0	2.69
job	category	200	1	0.5	0	0.0	0	0.0	retired	200
day_of_week	category	200	5	2.5	0	0.0	0	0.0	wed	65
month	category	200	1	0.5	0	0.0	0	0.0	may	200
contact	category	200	1	0.5	0	0.0	0	0.0	cellular	200
poutcome	category	200	1	0.5	0	0.0	0	0.0	nonexistent	200
loan	category	200	3	1.5	0	0.0	0	0.0	no	123
housing	category	200	3	1.5	0	0.0	0	0.0	yes	133
default	category	200	3	1.5	0	0.0	0	0.0	no	148
education	category	200	8	4.0	0	0.0	0	0.0	high.school	51
marital	category	200	4	2.0	0	0.0	0	0.0	married	91
y	category	200	2	1.0	0	0.0	0	0.0	False	164