Convolutional and recurrent neural networks, used in computer vision and natural language processing, respectively, have been shown to be effective at improving a variety of machine learning tasks. However, all of the inputs used by these deep learning paradigms are of the Euclidean structure type, such as text or images. Since graphs are a typical non-Euclidean structure in the machine learning area, it is challenging to directly apply these neural networks to graph-based applications like node classification. Due to increased research focus, graph neural networks—which are created to handle specific graph-based input—have made significant advancements. In this article, we present an in-depth review of the use of graph neural networks for the node classification task. The state-of-the-art techniques are first described and broken down into three primary groups: attention technique, convolutional technique, and autoencoder technique. The performance of several approaches is then compared in-depth comparative tests on a number of benchmark datasets.