Skip to main content
HomeTutorialsData Science

Synthetic Data Generation: A Hands-On Guide in Python

Learn everything you need to know about synthetic data generation. Discover the techniques and tools that make synthetic data essential for AI and machine learning with practical Python code examples to help you get started!
Oct 21, 2024  · 25 min read

In the post-GPT world, the demand for high-quality datasets has never been greater. Generative AI and large language models all hinge on the availability of robust, high-quality, and high-quantity data. However, obtaining such data is often faced with challenges, ranging from privacy concerns to the sheer scarcity of actual data worldwide. 

Enter synthetic data generation: a powerful solution that addresses the scarcity of actual data and settles privacy concerns. In this blog post, we will learn what synthetic data is, why it's essential, the techniques and tools for generating it, and how you can get started.

What is Synthetic Data?

Synthetic data is artificially generated information that mimics real-world data in structure and statistical properties but doesn't correspond to actual entities. It's created algorithmically and is used as a stand-in for real data in various applications.

The mentioned characteristics make synthetic data invaluable for software testing, AI model training, data augmentation, financial modeling, and healthcare research, where large, diverse datasets are needed, but real data may be scarce, sensitive, or restricted by privacy regulations.

An image by Gartner showing the growth of synthetic data in AI models over time (2020 - 2030)

Synthetic data generated from computer simulations or algorithms provides an inexpensive alternative to real-world data and is increasingly used to create accurate AI models. Image source: NVIDIA blog

Types of Synthetic Data

The three main types of synthetic data are structured, unstructured, and sequential, each with its own characteristics and applications. Let’s explore them in more detail.

1. Structured data

Structured data is information organized in a predefined format, typically found in databases and spreadsheets. This type of synthetic data mimics the format and statistical properties of real structured data. Examples include:

  • Customer records: Synthetic customer profiles with attributes like name, age, address, and purchase history.
  • Transaction logs: Artificially generated financial transactions or sales data.
  • Sensor readings: Simulated data from IoT devices or industrial sensors.
  • Healthcare records: Synthetic patient data, including diagnoses, treatments, and outcomes.

Generation methods for structured synthetic data often involve statistical modeling, rule-based systems, or machine learning techniques like variational autoencoders (VAEs) or generative adversarial networks (GANs).

2. Unstructured data 

Unstructured data lacks a predefined format and is increasingly important in machine learning applications. Types of synthetic unstructured data include:

3. Sequential data

Sequential data involves time or order, making it essential for models that need to understand temporal relationships. Examples include:

  • Time-series data: Synthetic stock prices, weather patterns, or energy consumption data.
  • Event logs: Artificially generated sequences of user actions on a website or app.
  • Sensor streams: Simulated continuous data from wearable devices or industrial equipment.

Become Data Science Certified

Supercharge your career as a professional data scientist.

Get Certified Today
Timeline mobile.png

Why Use Synthetic Data?

Synthetic data offers several advantages, making it an attractive option for organizations and researchers. Its use can address various challenges in data science, machine learning, and software development. Here are some of the reasons to use synthetic data:

1. Overcoming data scarcity

In many fields, collecting enough quality data is a significant hurdle. Synthetic data can fill these gaps, allowing for comprehensive analysis and robust model training.

  • Rare event simulation: Certain critical events occur infrequently in fields like finance or disaster management. Synthetic data can generate more examples of these rare events, improving model performance in these crucial scenarios.
  • Balanced datasets: Real-world data often suffer from class imbalance. Synthetic data can be generated to balance underrepresented classes, leading to more accurate models.
  • Supplementing small datasets: In domains where data collection is expensive or time-consuming, synthetic data can augment limited real data to reach the volume needed for effective machine learning.

2. Enhancing data privacy

By using synthetic data that doesn't correspond to actual individuals, organizations can sidestep privacy concerns and compliance with regulations like GDPR and HIPAA.

  • Anonymization alternative: Unlike anonymization techniques that can sometimes be reversed, synthetic data is generated from scratch, eliminating the risk of re-identification.
  • Cross-border data sharing: Synthetic data can facilitate international collaboration on sensitive projects without violating data protection laws.
  • Testing and development: Developers can work with realistic data without exposure to actual personal information, reducing the risk of data breaches during the development process.

3. Better machine learning models

Synthetic data can be tailored to include rare events or specific conditions, enriching datasets and improving model accuracy.

  • Edge case representation: Synthetic data can be generated to include edge cases crucial for model performance but rare in real data.
  • Bias reduction: By carefully designing synthetic data generation processes, it's possible to create datasets that are more diverse and representative than available real data, helping to reduce bias in AI models.
  • Transfer learning: Synthetic data can be used to pre-train models on tasks where real data is scarce, improving performance when fine-tuned on limited real data.

4. Cost efficiency

Synthetic data can often be more cost-effective than collecting or purchasing real data.

  • Reduced data acquisition costs: Generating synthetic data can be cheaper than collecting real data, especially for large-scale or specialized datasets.
  • Lower storage and maintenance costs: Synthetic data can be generated on demand, potentially reducing long-term storage costs associated with large real datasets.
  • Minimized legal and compliance expenses: By using synthetic data, organizations can reduce costs associated with data privacy compliance and potential legal issues.

Techniques for Synthetic Data Generation

Several methods exist for generating synthetic data, each with advantages and use cases. Let’s review them in this section. 

1. Random data generation

This is the simplest method, involving generating data by randomly sampling from statistical distributions.

import pandas as pd
import numpy as np

# Generate 1,000 samples from a normal distribution
heights = np.random.normal(loc=170, scale=10, size=1000)
weights = np.random.normal(loc=70, scale=15, size=1000)

# Create a correlation between height and weight
weights += (heights - 170) * 0.5
df = pd.DataFrame({'Height': heights, 'Weight': weights})

# Add categorical data
genders = np.random.choice(['Male', 'Female'], size=1000)
df['Gender'] = genders
df.head()
df.describe()

The above code generates a random dataset simulating height and weight using normal distributions. The heights are centered around 170 cm with a standard deviation of 10, and the weights are centered around 70 kg with a standard deviation of 15. A linear relationship is introduced between height and weight, so taller individuals generally weigh more. The code also adds a Gender column with randomly assigned values of “Male” or “Female.”

Output:

This method is quick and easy to implement, making it useful for basic testing or prototyping. However, it has some limitations. It doesn't capture complex relationships or real-world patterns or account for specific domain-related nuances or constraints.

2. Rule-based generation

Rule-based methods use predefined rules or domain knowledge to create data with specific characteristics. Take a look at the following example:

import numpy as np
import pandas as pd

def generate_customer_data(num_records):
    data = []
    for _ in range(num_records):
        age = np.random.randint(18, 80)
       
        # Rule: Income is loosely based on age
        base_income = 20000 + (age - 18) * 1000
        income = np.random.normal(base_income, base_income * 0.2)
       
        # Rule: Credit score is influenced by age and income
        credit_score = min(850, max(300, int(600 + (age/80)*100 + (income/100000)*100 + np.random.normal(0, 50))))
       
        # Rule: Loan amount is based on income and credit score
        max_loan = income * (credit_score / 600)
        loan_amount = np.random.uniform(0, max_loan)
       
        data.append([age, income, credit_score, loan_amount])
   
    return pd.DataFrame(data, columns=['Age', 'Income', 'CreditScore', 'LoanAmount'])

df = generate_customer_data(1000)
df.head()

The Python code above generates a dataset of customer data with 1,000 records. For each customer, it randomly assigns an age between 18 and 80. Based on age, an income is calculated, where older individuals generally have higher incomes. The credit score is then determined, influenced by age and income, ensuring it falls between 300 and 850. A loan amount is also generated, dependent on the individual's income and credit score. 

Output:

An image showing dataframe sample data with 5 rows

The used approach ensures that the generated data adheres to known relationships and constraints, making it possible to incorporate domain-specific knowledge directly into the data generation process. However, it can become complex when dealing with systems with many interrelated rules, and defining and maintaining such rules can be time-consuming for large or intricate datasets.

3. Simulation-based generation

Simulation-based data generation models real-world processes to create synthetic datasets. It involves simulating events or behaviors based on predefined rules and variables, allowing us to study how different factors interact in dynamic systems. 

This method is often used to test scenarios and analyze potential outcomes, providing valuable insights into complex systems like customer service processes, supply chains, and healthcare operations. Here’s an example in Python:

import simpy
import random
import pandas as pd

class Bank(object):
    def __init__(self, env, num_tellers):
        self.env = env
        self.teller = simpy.Resource(env, num_tellers)
       
    def service(self, customer):
        service_time = random.expovariate(1/10)  # Avg service time of 10 minutes
        yield self.env.timeout(service_time)

def customer(env, name, bank, data):
    arrival_time = env.now
    print(f'{name} arrives at the bank at {arrival_time:.2f}')
   
    with bank.teller.request() as request:
        yield request
        wait_time = env.now - arrival_time
        print(f'{name} waits for {wait_time:.2f}')
       
        yield env.process(bank.service(name))
       
        service_time = env.now - arrival_time
        print(f'{name} leaves the bank at {env.now:.2f}')
   
    # Append the customer data to the list
    data.append((name, arrival_time, wait_time, service_time))

def run_simulation(env, num_customers, bank, data):
    for i in range(num_customers):
        env.process(customer(env, f'Customer {i}', bank, data))
        yield env.timeout(random.expovariate(1/5))  # New customer every 5 minutes on average

# Set up and run the simulation
env = simpy.Environment()
bank = Bank(env, num_tellers=3)
num_customers = 100
data = []  # This will hold the customer data
env.process(run_simulation(env, num_customers, bank, data))
env.run()

# Create a DataFrame from the collected data
df = pd.DataFrame(data, columns=['Customer', 'ArrivalTime', 'WaitTime', 'ServiceTime'])
print(df.head())
print(df.describe())

The code simulates a bank using the SimPy library. Customers arrive at random intervals, wait for service from one of the three available tellers, and then leave after being served. 

The Bank class defines the service time (averaging 10 minutes), while the customer() function tracks each customer’s arrival, waiting time, and service duration. Customers are processed in the run_simulation() function, which generates new arrivals on average every 5 minutes.

Output:

An image showing dataframe sample data with 5 rows

This approach allows for modeling complex and dynamic systems, enabling scenario testing and "what-if" analyses. However, it can become computationally intensive for large-scale simulations and requires a thorough understanding of the modeled system to ensure accurate results.

4. Generative models

Generative models like generative adversarial networks (GANs) and variational autoencoders (VAEs) have become popular for creating highly realistic synthetic data. These models learn to capture the underlying distribution of the input data and generate new samples that mimic real-world data.

Both methods are widely used in tasks like image generation and natural language processing, primarily for creating synthetic data for machine learning models when real-world data is scarce or sensitive.

Check out the free course on Generative AI Concepts on Datacamp to learn how generative AI models are developed and how they will impact society in the future.

Tools for Synthetic Data Generation

Several tools are available that make it easier to generate synthetic data for various use cases, ranging from statistical modeling to machine learning and healthcare simulations. 

These tools are designed to create data that mirrors the characteristics of real-world datasets, allowing for safe experimentation and testing without exposing sensitive information.

1. Synthetic Data Vault (SDV) in Python 

SDV is a Python library that provides a suite of models for generating synthetic data. It supports various data types, including time series, relational data, and tabular data. SDV uses advanced probabilistic models like Gaussian copulas and deep learning methods to learn the structure of the original data and produce high-quality synthetic versions. It’s widely used in machine learning applications for generating synthetic training data when real data is limited or sensitive.

To learn more about generating synthetic data using the SDV library in Python, check out the official GitHub repository or try their official tutorials.

2. Gretel.AI

Gretel.AI is a cloud-based platform that offers synthetic data generation and anonymization services. It provides user-friendly APIs and pre-built models that can generate realistic synthetic data based on a variety of inputs, including structured, semi-structured, and unstructured data. Gretel.AI focuses on simplifying the workflow for developers and data scientists, making it easier to generate synthetic data with minimal setup.

Check out Gretel’s quick start to get started with synthetic data generation using Gretel.AI

3. Synthea

Synthea is an open-source tool specifically developed for healthcare simulations. It generates synthetic patient records that reflect realistic medical histories, treatments, and outcomes, following medical and clinical standards. Researchers and healthcare institutions often use Synthea to simulate patient populations and study public health trends without using real patient data, making it invaluable for healthcare data analysis.

4. synthpop (R programming language)

synthpop is an R package specifically designed to generate synthetic versions of real-world datasets. It works by using statistical models to replicate the relationships and distributions found in the original data, allowing users to safely share and analyze data without compromising confidentiality. synthpop is particularly useful for researchers and institutions that need to release datasets while protecting privacy.

To learn more about generating synthetic data in R using synthpop, check out the Generating Synthetic Data Sets with synthpop in R tutorial.

The Role of Generative AI in Synthetic Data

As previously mentioned, generative AI plays a significant role in creating synthetic data, mainly through advanced models that learn real-world data's underlying patterns and distributions and generate realistic synthetic versions. Let’s explore each of these techniques and their applications.

Generative adversarial networks (GANs)

GANs are one of the most powerful tools for generating synthetic data. They consist of two competing neural networks: a generator and a discriminator. The generator creates synthetic data, while the discriminator attempts to differentiate between real and synthetic data. Through this adversarial process, the generator improves its ability to create highly realistic data. GANs are widely used to generate images, videos, and even tabular data, and they have found applications in fields such as healthcare, autonomous driving, and entertainment.

Variational autoencoders (VAEs)

VAEs are another popular generative model for synthetic data generation. They work by encoding data into a compressed latent space and then decoding it back to its original form. This allows VAEs to generate new data points by sampling from the latent space, making them useful for applications like image synthesis and anomaly detection. While VAEs don’t achieve the same level of realism as GANs, they are more stable during training and provide better control over the structure of the generated data.

Transformer models for text generation

Transformers, like GPT models, have revolutionized text generation by producing coherent and contextually accurate sentences and paragraphs. For synthetic data, transformers can generate human-like text that mirrors real-world datasets, such as customer reviews, conversations, or other textual data. These models are particularly useful when generating large volumes of text-based synthetic data for natural language processing (NLP) tasks, but they also have the flexibility to be fine-tuned for specific applications.

To learn the role generative artificial intelligence plays today and will play in the future in a business environment, check out the Generative AI for Business course.

Getting Started with Synthetic Data Generation

1. Choosing the right technique

The appropriate synthetic data generation technique depends on your use case, data type, and goals. For simple datasets with well-defined statistical distributions, random data generation or rule-based generation can be sufficient. 

If you're dealing with complex, interrelated data or need to preserve relational integrity (e.g., in healthcare or finance), a simulation-based approach or generative models like GANs and VAEs might be more appropriate. Transformer-based models can provide better results for text generation or tasks involving large-scale data. 

Consider the trade-offs between ease of implementation, data realism, and computational requirements when choosing a technique.

2. Implementation in Python using Gretel.AI API 

To demonstrate this example, I have generated a mock patient database of three tables, each with exactly 5,000 records.

Table 1 - Patients

An image showing dataframe sample data with 5 rows

Table 2 - Lab Results

An image showing dataframe sample data with 5 rows

Table 3 - Treatments

An image showing dataframe sample data with 5 rows

To implement this example, you will have to sign up on Gretel.AI and generate an API key. You will also have to install a library in Python:

!pip install -U gretel-trainer

To implement this use case, we will use RelationalData from the gretel_trainer.relational module. It allows users to automatically define and maintain relationships between multiple tables, ensuring that foreign key dependencies and data integrity are preserved in the synthetic output. 

This is particularly useful for generating high-quality synthetic data that mirrors the structure and dependencies of real-world databases, making it easier to use in applications like machine learning, research, or compliance with data privacy regulations. 

Code Example:

import pandas as pd
from gretel_client import configure_session
from gretel_trainer.relational import RelationalData

configure_session(api_key="your_api_key", cache="yes", validate=True)

csv_dir = "csv/"
tables = [("patients", "Patient ID"), ("lab_results", "Lab Result ID"), ("treatments", "Treatment ID")]
foreign_keys = [("lab_results.Patient ID", "patients.Patient ID"), ("treatments.Patient ID", "patients.Patient ID")]

relational_data = RelationalData()
for table, pk in tables:
    relational_data.add_table(name=table, primary_key=pk, data=pd.read_csv(f"{csv_dir}/{table}.csv"))
for fk, ref in foreign_keys:
    relational_data.add_foreign_key(foreign_key=fk, referencing=ref)

from gretel_trainer.relational import MultiTable

multitable = MultiTable(
    relational_data,
    project_display_name="Clinical Trials",
    gretel_model="amplify"
)
multitable.train()

multitable.generate(record_size_ratio=1)
synthetic_patients = multitable.synthetic_output_tables['patients']
synthetic_lab_results = multitable.synthetic_output_tables['lab_results']
synthetic_treatments = multitable.synthetic_output_tables['treatments']

This code uses the Gretel.AI API to generate synthetic data for a relational database. It first sets up a session using an API key and loads three CSV files (patients, lab_results, and treatments), specifying their primary keys and defining foreign key relationships. 

The RelationalData class is used to maintain these relationships. After configuring the data, a model is trained using the MultiTable class, which synthesizes the entire relational dataset. 

After the model is trained, the generate() method is used to generate synthetic data. The record_size_ratio parameter determines the ratio reference to size of the generated data. record_size_ratio=1 means the number of records in synthetic dataset = original dataset i.e. 5,000 records.

3. Evaluating the quality of the synthetic dataset

Evaluating the quality of synthetic data is essential to ensure its utility and privacy. Start by comparing statistical distributions between the original and synthetic data—metrics like mean, variance, and correlation should closely match. 

Another method is to use synthetic data to train machine learning models and compare their performance (e.g., accuracy or F1 score) to models trained on real data. 

A visualization comparing Age Distribution of original data with synthetic data.

Example visualization to assess the quality of synthetic data. Image by Author.

Conclusion

Synthetic data generation has emerged to address challenges related to data scarcity, privacy concerns, and the need for diverse, high-quality datasets. 

With techniques ranging from simple random data generation to advanced generative models like GANs and VAEs, synthetic data enables organizations to innovate and build better machine learning models while safeguarding sensitive information. 

Tools like SDV and managed services like Gretel.AI make generating synthetic data more accessible, even for complex relational databases. 

As synthetic data generation evolves, its applications will be increasingly important in advancing AI, healthcare, finance, and countless other fields.

Want to dive deeper? Watch the Using Synthetic Data for Machine Learning & AI in Python training to discover what synthetic data is, how it protects privacy, and how it's being used to accelerate AI adoption in banking, healthcare, and many other industries.

Get certified in your dream Data Scientist role

Our certification programs help you stand out and prove your skills are job-ready to potential employers.

Timeline mobile.png

FAQs

How does synthetic data differ from real data?

Synthetic data is artificially generated to mimic the statistical properties of real data but does not represent actual entities. While real data is collected from real-world sources, synthetic data is produced through algorithms, making it ideal for overcoming privacy and data scarcity challenges.

Is synthetic data as reliable as real data for training AI models?

Yes, synthetic data can be just as reliable, provided it's generated with high-quality methods that accurately capture the patterns and relationships in real data. In some cases, it can even enhance models by introducing rare events or edge cases not typically found in real datasets.

Can I use synthetic data for regulatory compliance?

Yes, synthetic data can help with regulatory compliance by eliminating privacy concerns since it does not correspond to real individuals or sensitive information. This makes it useful for testing and sharing data across borders without violating data protection laws.

Are there any limitations to using synthetic data?

While synthetic data is a powerful tool, it may not capture the full complexity of real-world data, especially in highly nuanced domains. Additionally, synthetic data may introduce bias or unrealistic patterns if the generation process is not carefully designed. Therefore, it's important to evaluate the limitations of synthetic data in your specific use case before relying on it exclusively.


Photo of Moez Ali
Author
Moez Ali
LinkedIn
Twitter

Data Scientist, Founder & Creator of PyCaret

Topics

Learn more about machine learning and AI with the following courses!

Course

Understanding Machine Learning

2 hr
209.8K
An introduction to machine learning with no coding involved.
See DetailsRight Arrow
Start Course
See MoreRight Arrow
Related

blog

What is Synthetic Data?

Synthetic data is artificially generated data that mimics the characteristics of real-world data without containing any actual information.
Abid Ali Awan's photo

Abid Ali Awan

6 min

tutorial

Creating Synthetic Data with Python Faker Tutorial

Generating synthetic data using Python Faker to supplement real-world data for application testing and data privacy.
Abid Ali Awan's photo

Abid Ali Awan

13 min

tutorial

A Complete Guide to Data Augmentation

Learn about data augmentation techniques, applications, and tools with a TensorFlow and Keras tutorial.
Abid Ali Awan's photo

Abid Ali Awan

15 min

tutorial

Python Tutorial for Beginners

Get a step-by-step guide on how to install Python and use it for basic data science functions.
Matthew Przybyla's photo

Matthew Przybyla

12 min

code-along

Using Synthetic Data for Machine Learning & AI in Python

Rewatch this training to discover what synthetic data is, how it protects privacy, and how it's being used to accelerate AI adoption in banking, healthcare, and many other industries.
Alexandra Ebert's photo

Alexandra Ebert

code-along

Getting Started with Machine Learning in Python

Learn the fundamentals of supervised learning by using scikit-learn.
George Boorman's photo

George Boorman

See MoreSee More