GDEGAN: Graphical Discriminative Embedding GAN for tabular data

While generative models achieve remarkable success in recent years, applying them to model tabular data is still challenging. The first problem of tabular data is the categorical encoding scheme, in which each categorical value is represented as a one-hot vector. It leads to the problem of very sparse high dimensional input space, especially when the cardinality of values of each attribute is high. This is problematic to GAN training since a trivial discriminator can simply distinguish real and fake data by checking the distributions sparseness. The second problem in tabular data is its hard constraint and discrete signals property which is challenging for neural networks. The modelling of discrete features is often associated with counting problem where gradient signals are not well-prepared for. As a result, current GAN methods might overlook and ignore these discrete features causing the mode-collapse in tabular data modelling. In this paper, we propose a unified framework to solve these two problems: (i) we propose to embed the raw data into the highlevel features and train GAN these features instead to avoid the trivial sparseness detection by the discriminator (ii) we propose the graphical-conditional vector to encourage GAN to learn to generate the structure information across multiple attributes. The experimental results show that our proposed methods perform much better than the current state-of-the-art method on most tabular benchmark datasets. Source code is available at: https://github.com/dungdinhanh/GDEGAN