Epilepsy diagnostic investigation involving manual visual analysis of electroencephalogram (EEG) is a time-consuming process. Deep neural networks, especially the convolutional network (CNN), have been applied to interictal epileptiform discharge (IED) detection and have achieved promising results. However, these networks do not incorporate clinical features of EEG montages. In recent years, graph convolution has succeeded in learning features from structural graph-like data. In this paper, we explore the novel application of different architectures of graph convolutions with Chebyshev polynomial filters which learn spatio-temporal features from EEG montages. We conducted a number of experiments with transverse and longitudinal montages on a set of routine EEG recordings from patients with idiopathic generalized epilepsy. We split these EEG recordings into 2s windows with or without IED and evaluated different architectures in terms of how well they classified these windows. We achieved the best AUC of 0.92. Furthermore, we explored different thresholds of the output probability and observed that at 0.8, based on the selection of collected data, we achieved a mean false-positive rate per minute of 0.44 and still preserved a reasonable mean sensitivity of 0.64 across all EEG recordings. The results indicate that our approaches could produce clinically useful performance levels. Our work could be extended to improve the interpretability of the automated software in a clinical environment.