Synthetic Data: Dataset generation for your need

By Aditya Abeysinghe

During the last few decades, research into various fields has expanded rapidly with advances in technology and tools used for research. In many fields, data is expensive to be generated or found from sources matching research requirements and sometimes is inaccessible due to privacy concerns or sometimes data is unavailable. In these kinds of cases, data needs to be artificially generated to match real world data and tested. These kind of data is called synthetic data and is now considered an emerging topic in many research fields.

Synthetic data is not a new model; it has its origins since 1980s when researches started to grow with advances in technologies. However, it became an emerging topic with emerging technologies such as self-driving cars where data required for training using algorithms is hard to be found. Meanwhile, many restrictions on data sharing cause data in many locations to be restricted to be shared with other locations.

Is synthetic data comparable to real world data?

There have been many studies that have evaluated the performance of synthetic data with that of real world data. These types of research found that nearly 70% of the time synthetic data show the same results as those shown by real world data. This was compared by use of both kinds of data in machine learning algorithms and then comparing the performance by both.

In fields such as data analytics, there have been many doubts on the use of synthetic data as these data have been not verified by real world experiments. Therefore, analysts are often reluctant to carry out analysis in many fields on synthetic data. Meanwhile, in fields such as data science, synthetic data is often used as a secondary source when primary data is often unavailable. In many emerging technologies, synthetic data is the norm where research on these technologies is limited.

Benefits of using synthetic data

The main benefit of using synthetic data is that the user can generate the dataset as per their requirement. Datasets commonly available often have attributes that are not required by users, have uncleaned data and the format of the data is often not suited to be trained by algorithms. When data is synthetic, the user can generate only the required attributes and number of records where format, empty values and redundant values need not be thought during model building. In other words, the dataset is easily customizable to the needs of the user.

Another benefit of generating the dataset synthetically is preserving the privacy of data. Privacy enhancing computation is an emerging field, as described in my previous article. Privacy enhancing computation is required when privacy needs to be ensured during data sharing. Several methods of privacy enhancing computation exist. However, if the data is generated synthetically, data could be generated preserving privacy. Then the need for later privacy enhancing methodologies is often eliminated.

Image Courtsey: https://internationaljournalofresearch.com/