Feature Processing on Artificial Graph Node Features for Classification with Graph Neural Networks

Graph Neural Networks are often used for graph-related problems such as graph, node, and edge classification. Node features and connections between nodes are used for learning graph representations for such classification problems. However, many graph datasets don’t have node attributes and this leads to an obstacle while working with featureless graph datasets. It’s possible to overcome this obstacle with various feature initialization methods. These methods can be categorized as structural and positional artificial feature initialization methods. Features that are generated by these methods are not always flawless. Sometimes these features can contain unnecessary information that leads to decreased classification performance and increased memory consumption and training time. Based on our research it has been seen that there are very few studies addressing this problem. In this study, we aim to overcome this problem by using structural graph information. We propose a method that is mainly focused on computing measures of nodes and creating node representations by clustering model. Our proposed method computes node measures such as degree, clustering coefficient, number of triangles, and average neighbor degree and trains a clustering model. Experiments on datasets such as Malnet, Github Stargazers, MUTAG, and DD graph datasets have shown that our proposed method which is based on node measures and clustering improves the classification performance compared to artificial node degree features and achieves competitive performance against the DeepWalk method.