Survey on Synthetic Data Generation, Evaluation Methods and GANs

Abstract

Synthetic data consists of artificially generated data. When data are scarce, or of poor quality, synthetic data can be used, for example, to improve the performance of machine learning models. Generative adversarial networks (GANs) are a state-of-the-art deep generative models that can generate novel synthetic samples that follow the underlying data distribution of the original dataset. Reviews on synthetic data generation and on GANs have already been written. However, none in the relevant literature, to the best of our knowledge, has explicitly combined these two topics. This survey aims to fill this gap and provide useful material to new researchers in this field. That is, we aim to provide a survey that combines synthetic data generation and GANs, and that can act as a good and strong starting point for new researchers in the field, so that they have a general overview of the key contributions and useful references. We have conducted a review of the state-of-the-art by querying four major databases: Web of Sciences (WoS), Scopus, IEEE Xplore, and ACM Digital Library. This allowed us to gain insights into the most relevant authors, the most relevant scientific journals in the area, the most cited papers, the most significant research areas, the most important institutions, and the most relevant GAN architectures. GANs were thoroughly reviewed, as well as their most common training problems, their most important breakthroughs, and a focus on GAN architectures for tabular data. Further, the main algorithms for generating synthetic data, their applications and our thoughts on these methods are also expressed. Finally, we reviewed the main techniques for evaluating the quality of synthetic data (especially tabular data) and provided a schematic overview of the information presented in this paper.

Scaling While Privacy Preserving: A Comprehensive Synthetic Tabular Data Generation and Evaluation in Learning Analytics

“…Once the training was finished, we used the data generation models to generate synthetic tabular data of 5000 rows. The rationale for this size of the synthetic datasets is twofold: i) we aimed at a computationally feasible data set size, which would allow us to examine the effect of different ratios of real to synthetic data on the evaluation measures, ii) synthetic dataset of such size was used in an earlier similar study [18]. The synthetic dataset and the original training dataset were then separately used to train the above-mentioned three machine learning algorithms and the resulting models were evaluated on the original test dataset.…”

Evaluating Synthetic Medical Images Using Artificial Intelligence with the GAN Algorithm

“…GANs have undergone several modifications since they were first proposed to solve several different problems in different domains, e.g., physics [ 18 ], healthcare [ 19 ], or object detection [ 20 ]. To analyze the state-of-the-art in what concerns GANs used for synthetic data generation, as well as synthetic data generation methods, we reviewed recently published scientific papers [ 21 , 22 , 23 ]. Pose-driven attention-guided image generation for person re-Identification proposed in [ 24 ] by Amena et al introduces attentive learning and transferring the subject pose through an attention mechanism based on GAN.…”

Can We Trust Synthetic Data in Medicine? A Scoping Review of Privacy and Utility Metrics

“…A few surveys in the field have examined various aspects of synthetic data generation 19,20 . Figueira et al 19 provide an extensive description of multiple generation methods while Hernandez et al 20 explored evaluation methods and compared them to determine the best-performing ones.…”

“…A few surveys in the field have examined various aspects of synthetic data generation 19,20 . Figueira et al 19 provide an extensive description of multiple generation methods while Hernandez et al 20 explored evaluation methods and compared them to determine the best-performing ones. In contrast to these prior studies, our approach differs in how we identify the obstacles hindering the adoption of synthetic data as we place a greater emphasis on the evaluation process and the privacy-utility trade-off dilemma by having a systematic look at how synthetic data is evaluated across 92 studies.…”