Feature-Aware Unsupervised Detection of Important Nodes in Graphs

Mohammadreza Ghanbari, Saeed Asadi Bagloee, Jianzhong Qi, Majid Sarvi

Abstract: Identifying important nodes in a graph is a crucial task in many applications such as biological networks, transportation networks, and social networks. The main difficulty in this task is the lack of ground truth. Existing methods mainly focus on the structure of the graphs and ignore the node features or operate in a supervised manner. In this paper, we design a trainable model named FadiGNN that incorporates node feature information as well as topology structure in an unsupervised manner to score the nodes within a graph. It consists of Graph Convolutional Networks (GCNs) and adapted personalized PageRank, along with a loss function that ensures convergence of the model. To show the effectiveness of FadiGNN, experiments are conducted on two different application settings; node classification and active learning. The experiments are conducted on different real-world datasets including large graph data. The results on node classification shows that FadiGNN outperforms the state-of-the-art models by an average of 7.62% in terms of accuracy.