Graph neural networks in node classification: survey and evaluation

Neural networks have been proved efficient in improving many machine learning tasks such as convolutional neural networks and recurrent neural networks for computer vision and natural language processing, respectively. However, the inputs of these deep learning paradigms all belong to the type of Euclidean structure, e.g., images or texts. It is difficult to directly apply these neural networks to graph-based applications such as node classification since graph is a typical non-Euclidean structure in machine learning domain. Graph neural networks are designed to deal with the particular graph-based input and have received great developments because of more and more research attention. In this paper, we provide a comprehensive review about applying graph neural networks to the node classification task. First, the state-of-the-art methods are discussed and divided into three main categories: convolutional mechanism, attention mechanism and autoencoder mechanism. Afterward, extensive comparative experiments are conducted on several benchmark datasets, including citation networks and co-author networks, to compare the performance of different methods with diverse evaluation metrics. Finally, several suggestions are provided for future research based on the experimental results.