A Message Passing Neural Network Framework with Learnable PageRank for Author Impact Assessment

Abstract

The assessment of author influence is crucial for the advancement of scientific research and policy shaping in academia. PageRank and its derivatives, primarily focusing on network topology, often overlook spatial attributes and exhibit biases, besides being inefficient due to their iterative nature. We propose a novel Neuro-Enhanced PageRank Network (NPRNet), which integrates graph neural networks with PageRank to address these deficiencies. NPRNet utilizes Message Passing Neural Networks to efficiently compute and incorporate learnable parameters, thus considering node attributes. A semi-supervised learning strategy is also developed to manage the absence of true labels. Validated using conference articles in the field of artificial intelligence (AI) from Scopus API since 1985, NPRNet not only enhances computational efficiency but also effectively captures both topological and spatial feature information. It identifies leading countries in AI research, closely aligning with global trends in AI innovation and demonstrating the capacity to recognize recently active authors. This highlights its ability to reflect current research dynamics, thus deepening evaluations by integrating node attributes and supporting advanced knowledge management in research.

Keywords

• bibliometrics
• data mining
• neural networks
• network topology
• semisupervised learning