Indirect taxation is a significant source of income for any nation. Tax evasion hinders the progress of a nation. It causes a substantial loss to the revenue of a country. We design a model based on variational graph autoencoders and clustering to identify taxpayers who are evading indirect tax by providing false information in their tax returns. We derive six correlation parameters (features) and three ratio parameters from the data submitted by taxpayers in their returns. We derive four latent features from these nine features using variational graph autoencoder and cluster taxpayers using these four latent features. We identify taxpayers located at the boundary of each cluster by using kernel density estimation, which is further investigated to single out tax evaders. We applied our method to the iron and steel taxpayers data set provided by the Commercial Taxes Department, the government of Telangana, India. © 2021, Springer Nature Switzerland AG.