Bi-discriminator GAN for tabular data synthesis

Highlights

• Developing a novel data preprocessing scheme using Chi-squared function.

• Proposing a new conditional term for the generator network in a GAN setup.

• Implementing a bi-discriminator GAN for stable training.

• Designing straightforward architectures for generator and discriminator networks.

Abstract

This paper introduces a bi-discriminator GAN for synthesizing tabular datasets containing continuous, binary, and discrete columns. Our proposed approach employs an adapted preprocessing scheme and a novel conditional term using the ${\chi }_{\beta }^2$ distribution for the generator network to more effectively capture the input sample distributions. Additionally, we implement straightforward yet effective architectures for discriminator networks aiming at providing more discriminative gradient information to the generator. Our experimental results on four benchmarking public datasets corroborates the superior performance of our GAN both in terms of likelihood fitness metric and machine learning efficacy.

Introduction

Tabular data is among the most common modalities which has been widely used for maintaining massive databases of financial institutions, insurance corporations, networking companies, healthcare industries, etc. [1], [2], [3], [4], [5]. These databases include immense combination of personal, confidential, and general records for every customer, client, and patient in different formats (e.g., continuous and discrete data types). Semantic patterns derivable in such records efficiently contribute to extract meaningful information for the benefit of companies in various aspects such as large-scale decision-making [6], risk management [7], long-term investment [8], fraud or unusual activity detection [9], etc. However, exploiting these patterns is a challenging task since tabular datasets are heterogeneous [10], [11] and they contain sparse representations of discrete and continuous records with low correlation compared to homogeneous datasets (e.g., speech, environmental audio, image, etc.) [12]. Unfortunately, extracting semantic relational patterns from heterogeneous datasets requires implementing costly data-driven algorithms [13], [14].

During the last decade and especially after the proliferation of deep learning (DL) algorithms, various cutting-edge approaches have been introduced for processing tabular datasets in different frameworks [15], [16], [17], particularly for synthesis purposes [18]. Presumably, this is due to two major applications. Firstly, complex DL algorithms configured in the generative adversarial network (GAN) [19] synthesis platforms can be used for augmenting sparse datasets with low cardinality and poor sample quality [20]. Often, this data augmentation procedure effectively triggers the semantic pattern extraction operations. Secondly, GAN-based synthesis approaches yield models capable of generating new records similar (and non-identical) to the ground-truth samples available in the original databases. The synthesized records can be used for development purposes such as extracting relational patterns [21] without publicly releasing the original dataset. This efficiently contributes to protect the privacy of people and clients whose their information is stored in the tabular datasets of companies. Our focus in this paper is on the latter application since it has been among the most demanding appeals of some large-scale financial institutions towards avoiding data leakage [22], [23]. Briefly, we make the following contributions in this paper:

• developing a novel data preprocessing scheme and defining a new conditional term (vector) for the generator network configured in a GAN synthesis setup,

• implementing a bi-discriminator GAN for providing more gradient information to the generator network in order to improve its performance in runtime,

• designing straightforward architectures for generator and discriminator networks.

The organization of this paper is as the following. Section 2 provides a summary of synthesis approaches based on the state-of-the-art GANs for tabular datasets. In Section 3, we explain the details of our proposed synthesis approach and finally in Section 4, we report and analyze our conducted experiments on four benchmarking databases.