The Pros and Cons of Test Data Synthetics (or Data Fabrication)
by Louay Hazami
Data privacy is one of the most pressing issues in the new digital era. Data holds so much value for normal internet users and for all types of companies that are looking to capitalize on this new resource. To keep data anonymous and private, two types of solutions present themselves: masked data or synthetic data.
In the industry, synthetic data is more widely used. That’s because it’s much more secure than masked data. Therefore, I will be focusing on synthetic data. However, we can further our learning about synthetic data by comparing it with masked data.
So, in this post, I will be discussing the pros and cons of creating a synthetic dataset for testing purposes. First, I will define synthetic data by comparing it with masked data. I will also talk about how we can create synthetic data. Finally, we’ll discuss the challenges and benefits of using synthetic data.
Synthetic vs. Masked Data
First, let me explain each type of data:
- Masked data: Modified data that has a similar structure to the real data. Generally, we anonymize only sensitive data from the original data. It could be something as simple as changing the variable name. In contrast with synthetic data, we can keep some of the original data in the end product.
- Synthetic data: Data that is artificially manufactured from the original data. It has the same statistical properties. Creating synthetic data keeps all of the original data anonymous and unidentifiable. We can use all sorts of algorithms to map the synthetic data to the original data. Unlike with masked data, when we only give the synthetic data to the user, they have no idea what data they are dealing with most of the time. But because the synthetic data has the same statistical properties as the original data, the user can still use it to reach relevant conclusions.
For example, let’s say we have a database containing the credit card numbers of our customers. Using the anonymization process, we can simply hide most of the numbers of the credit card (571 is masked to become **1). Therefore, masked data still leaves some of the real information out there. However, we can use a machine learning algorithm to map the credit card numbers to another arbitrary number (571becomes 41273). And this is called synthetic data.
How to Create Synthetic Data
There are two main methods of creating synthetic data:
- Distribution-based modeling: This method relies on reproducing the statistical properties of the original data. For example, we can reproduce the variance or the mean of the data. Basically, we create new data points that have these same properties.
- Agent-based modeling: This method relies on creating a model. This model focuses on learning the behavior of the data algorithmically on its own. Depending on the data, this behavior can be simple or complex. It can also represent relationships between the different variables of the data. Then, the agent creates random data based on the observed properties.
In short, synthetic data can be created in two ways with varying levels of complexity. It may appear as a simple operation. I would like to present to you some of the tools used to create synthetic data next.
Tools to Create Synthetic Data
In addition to Enov8 DCS very own Data Fabrication module, there are multiple open-source resources available to create synthetic data. I will introduce three tools you can use for different methods of creating synthetic data:
- datasets: The Python package scikit-learn contains many tools that data scientists use. It also contains a module called datasets. In sklearn.datasets, we can find many methods for generating data samples. For example, a method for generating a dataset for a regression problem, make_regression, is available. Using make_regression, we can do agent-based modeling for example. The regressor here is our agent.
- pydbgen: This Python package allows us to generate random tables and databases. It also allows us to include specific types of data such as emails or dates in the tables.
- DCGANs: Deep Convolutional Generative Adversarial Networks (DCGANs) are neural networks capable of generating artificial images that look real. They are trained to use real images and output new images that look similar to the real images. Generally, this deep learning technique is used in data augmentation. We use data augmentation when we don’t have enough original images. So, we resort to using synthetic images to compensate for it.
Depending on the task, different techniques can be used. Generally, I think it’s not necessary to know how to use all of these methods. But knowing at least a few by heart can be helpful. Most importantly, I believe we need the ability to comprehend and study all the techniques whenever we encounter a problem where we need to use them.
Challenges of Using Synthetic Data
Undoubtedly, the ability to create a new set of data without revealing precious private information is very powerful. It offers us numerous possibilities for research and testing. However, we need to keep in mind something important: Not everything that works on the synthetic data will work on the original data. We can only study general trends with synthetic data. It’s very important to remember this during the testing phase. But there are other challenges as well:
- Users may lack trust in the data.
- If the quality of the data model is not high, we will reach the wrong conclusions.
- Complex data behavior can be very complex to replicate.
- If we simplify the representations in the data, the algorithms will not perform well.
- Synthetic data will require validation against the original data.
Because of these challenges, it can be difficult to create and use synthetic data to its full potential. We need to have a detailed strategy before diving in. For example, it’s good to plan to create different types of synthetic data for different use cases.
Benefits of Using Synthetic Data
However difficult it is to use synthetic data, it comes with benefits. Synthetic data becomes more important as the database grows. It helps us overcome the complexity of the data and privacy issues.
Firstly, synthetic data can replicate the trends of the original data. As a result, we can use it without breaking privacy rules. Moreover, we can use it to simulate new situations and conditions, such as rare weather or equipment malfunctioning scenarios. It also works well for prototype testing.
Furthermore, we can use synthetic data to combat overfitting. Overfitting happens when the algorithm performs well while training but fails during testing. In this case, we can use synthetic data to help train the algorithm. We can use it to make the data well balanced.
To sum up, synthetic data has multiple use cases. It also provides many solutions to real-world problems.
Synthetic Data as a Secret Recipe
After reading this post, do you feel that synthetic data will revolutionize your project? Certainly, it is exciting. We can solve many issues using it and improve our data models.
However, it’s important to rein in the expectations we have from synthetic data. More importantly, we need to understand how to use them and to what extent. Relying solely on synthetic data can make many algorithms basically useless in the long run. Companies should test on synthetic data, but they have to test on the original data before deploying their algorithms.
It is important to keep in mind that synthetic data is not 100% accurate. In short, it does not fully match the statistical properties of the original data. This is a persistent problem, but advancements are being made in this area.
In conclusion, I would like to say that synthetic data is an important tool in any data-related job. We should not overlook its benefits, but we should not glorify it. Like any solution, it has its limitations. And, as a data scientist, you should be aware of those limits. It’s important to communicate to management that they shouldn’t rely heavily on it because of privacy rules. They should know that the end goal of using synthetic data during testing is to get a good performance on the real-world data.
19 FFEBRUARY, 2021 by Carlos Schults "You can't improve what you don't measure." I'm sure you're familiar with at least some variation of this phrase. The saying, often attributed to Peter Drucker, speaks to the importance of metrics as fundamental tools to enrich and...
08 FEBRUARY, 2021 by Zulaikha Greer Data is the word of the 21st century. The demand for data analysis skills has skyrocketed in the past decade. There exists an abundance of data, mostly unstructured, paired with a lack of skilled professionals and effective tools to...
04 JANUARY, 2021 by Ukpai Ugochi Have you ever wondered what would happen if you mistakenly added bugs to your codes and shipped them to users? For instance, let's say an IT firm has its primary work tree on GitHub, and a team member pushes codes with bugs to the...
07 DECEMBER 2020 by Daniel de Oliveira In today’s application-based world, companies are releasing more applications than ever before. Software delivery life cycles are becoming more complicated. As a result, large companies require hundreds and even thousands of test...
19 NOVEMBER, 2020 by Michiel Mulders What Makes a Good Test Data Manager? Have you implemented test data management at your organization? It will surely benefit you if your organization processes critical or sensitive business data. The importance of test data is...
18 SEPTEMBER 2020 by Arnab Chowdhury Every aspect of our daily lives involves the usage of data. Be it our social media, banking account, or even while using an e-commerce site, we use data everywhere. This data may range from our names and contact information to our...