Graph neural networks for positive and unlabeled learning: a rewiring approach

Author(s)

Messias, Guilherme Henrique ; Iasulaitis, Sylvia ; Valejo, Alan Demetrius Baria

• Document type: Journal article
• Source: KNOWLEDGE AND INFORMATION SYSTEMS; v. N/A, p. 23-pg., 2025-08-01.

Abstract

Graph rewiring strategies have emerged as powerful tools to enhance vertex connection within a graph. These connections can be used by graph neural networks (GNNs) to improve classifier performance. However, rewiring methods are unsupervised and aim to improve connection metrics. This work introduces the positive sequential rewiring via breadth search (PSRB) model, an approach that applies supervised rewiring to the available classes in a positive and unlabeled learning (PUL) problem. PSRB employs a probability density function to select edge addition between positive vertices, facilitating the sequential application of a relational GNN. Experiments on five datasets from the literature show the algorithm's efficiency compared to state-of-the-art two-step PUL algorithms. (AU)

• FAPESP's process: 22/03090-0 - Analysis of large amounts of political data and complex networks: mining modelling and applications in Computational Political Science
• Grantee: Sylvia Iasulaitis
• Support Opportunities: Research Grants - Initial Project