DATGAN: Integrating expert knowledge into deep learning for synthetic tabular data

Abstract

Synthetic data can be used in various applications, such as correcting bias datasets or replacing scarce original data for simulation purposes. Generative Adversarial Networks (GANs) are considered stateof-the-art for developing generative models. However, these deep learning models are data-driven, and it is, thus, difficult to control the generation process. It can, therefore, lead to the following issues: lack of representativity in the generated data, the introduction of bias, and the possibility of overfitting the sample's noise. This article presents the Directed Acyclic Tabular GAN (DATGAN) to address these limitations by integrating expert knowledge in deep learning models for synthetic tabular data generation. This approach allows the interactions between variables to be specified explicitly using a Directed Acyclic Graph (DAG). The DAG is then converted to a network of modified Long Short-Term Memory (LSTM) cells to accept multiple inputs. Multiple DATGAN versions are systematically tested on multiple assessment metrics. We show that the best versions of the DATGAN outperform state-of-the-art generative models on multiple case studies. Finally, we show how the DAG can create hypothetical synthetic datasets.

Rigorous Experimental Analysis of Tabular Data Generated using TVAE and CTGAN

“…• DATGAN [31] proposes DATGAN which is a novel architecture based on GAN using Directed Acyclic Graphs (DAGs) to model the information about the dataset. It uses LSTM cells to model expert knowledge using DAG.…”

GAN-Based Tabular Data Generator for Constructing Synopsis in Approximate Query Processing: Challenges and Solutions

“…It is difficult for GAN to control the generation process of data-driven systems; therefore, integrating prior knowledge about data relationships and constraints can assist the generator in generating synopses that are realistic and meaningful. In order to implement this, DATGAN [38] incorporates expert knowledge into the GAN generator by matching the generator structure to the underlying data structure using a Directed Acyclic Graph (DAG). Using DAG, the nodes represent the columns of a data table, while the directed links between them allow the generator to determine the relationship between variables so that one column's generation influences another.…”

“…However, in AQP, it is not necessary to meet this threshold in order to generate realistic data synopses. DATGAN [38] uses the improved version of the Wasserstein loss function in WGAN [41] in addition to the Vanilla GAN loss function with a gradient penalty [42] and also adds the KL-divergence as an extra term to the original loss function. Both of these terms aim to minimize the difference between the probability distributions of real and generated data.…”

“…returns the input features of a given row. As mentioned before, DATGAN [38] uses DAG to control the generation process based on semantic relationships and correlations between columns. According to the constructed DAG, each column and its sequence are represented by Long Short-Term Memory (LSTM) cells.…”

CTTGAN: Traffic Data Synthesizing Scheme Based on Conditional GAN

“…They focus on the generation of high-dimensional discrete variables (binary and count features). Lederrey et al [ 34 ] proposed DATGAN model to generate population data. They combined expertise and deep learning methods and used directed acyclic graph to identify the relationships between variables.…”