Learning to rank complex network node based on the self-supervised graph convolution model

摘要

Node ranking in complex networks is of great value in many fields, such as identifying opinion leaders in social networks and risky institutions in financial networks. However, traditional ranking methods are built on heuristic rules and may be effective in certain datasets and fail in others. To address the issues, we formulate node ranking on complex networks as a learning to rank problem, and propose a novel model based on self-supervised learning and graph convolution model to rank nodes based on integrated information from node features and network structure. (1) We develop a self-supervised pretext task to extract information about node location and global topology from complex networks. (2) To train the model more efficiently, multi-task learning is adopted, which includes ranking task and regression task besides the self-supervised pretext task. Our model works well with only a small number of nodes that have ranking labels. Comprehensive experiments on different datasets demonstrate that it outperforms the existing state-of-the-art methods, which verify the effectiveness and robustness of the proposed approach.