Assessing citation appropriateness through masked graph autoencoder

This study proposes a novel network-based detection method for assessing citation reliability and identifying manipulative practices. Applying a Masked Graph Autoencoder, the approach learns expected citation patterns by iteratively masking and reconstructing network edges. High reconstruction errors flag manipulated citations, while genuine citations are accurately reconstructed due to their structural robustness. The reconstruction error provides a quantitative measure of reliability, enabling automated, large-scale detection of citation anomalies. This graph-based deep learning approach captures complex network relationships, reducing the need for manual audits.