Abstract
Academic citation networks are increasingly plagued by manipulative practices such as citation cartels and self-citation inflation. Traditional detection methods often fail to capture the complex structural and semantic nuances of these anomalies. This paper introduces GLAD (Graph Learning for Anomaly Detection), a multi-modal framework that combines textual analysis of citation contexts with graph neural networks to identify fraudulent edges. By leveraging node and edge feature learning alongside a specialized citation purpose classifier, GLAD achieves significant performance gains over state-of-the-art graph models.
Introduction
In the contemporary academic landscape, citation counts serve as a primary metric for research impact. This "academic currency" has incentivized various forms of manipulation, including citation cartels—groups of authors or journals that excessively cite one another—and coercive citations. Detecting these anomalous links is challenging because they often appear structurally similar to legitimate citations. Most existing Graph Neural Networks (GNNs) are designed for node-level anomaly detection, whereas citation fraud is inherently an edge-level problem. GLAD addresses this gap by treating citation detection as a binary classification task, focusing on the relationship between papers rather than just the papers themselves.
Methodology
The GLAD framework is built upon three core technical components designed to capture the identity, intent, and intensity of citations.
Node Feature Learning
To establish a robust representation of each paper (node), GLAD utilizes Deep Graph Infomax (DGI). This process fuses textual information—abstracts processed via Doc2vec—with structural importance metrics, such as in-degree and out-degree centrality. This ensures that the model understands both the content and the relative influence of each node within the network.
Citation Purpose (CPU) Algorithm
A critical innovation in GLAD is the CPU algorithm, which analyzes the semantic intent behind a citation. Using SciBERT, a language model pre-trained on large-scale scientific corpora, the algorithm classifies citation contexts into six categories: Criticizing, Comparison, Use, Substantiating, Basis, or Other. Citations lacking a clear scientific purpose or appearing in unrelated contexts are flagged as suspicious.
Edge Feature Learning
The model employs a deep autoencoder to learn representations of the links (edges) between papers. It incorporates metadata such as prior collaborations, institutional affiliations, and self-citation status. By integrating these node and edge representations into a joint loss function, GLAD can distinguish between legitimate scholarly discourse and "gamed" citations.
Experimental Setup
Due to the lack of standard benchmarks for citation fraud, a custom dataset was constructed:
- Network Structure: A citation network of 4,718 relationships derived from the Microsoft Academic Graph (MAG).
- Anomaly Simulation: Anomalies were identified using real-world cases of retracted papers and documented journal cartels.
- Annotation: Citation contexts were manually annotated by postgraduate students to provide ground truth for the CPU algorithm.
Results and Discussion
GLAD demonstrates superior performance compared to traditional baselines like DeepWalk, Node2vec, and GraphSAGE.
- Performance Metrics: The framework achieved an F1-score of 0.9075 and a precision of 0.9683. The high precision is particularly vital for academic integrity tools to minimize false accusations.
- Ablation Study: Removing either node or edge learning components resulted in a 20-40% drop in precision, confirming that a multi-modal approach is essential for capturing the complexity of citation fraud.
- Key Insight: The research highlights that citation counts are insufficient without considering citation purpose. A citation's validity is often found in its textual context rather than its mere existence.
Conclusion
GLAD provides a powerful tool for maintaining academic integrity by identifying anomalous citations that traditional metrics miss. Future work will focus on scaling the manual labeling process, incorporating more complex structural indices, and exploring temporal graph learning to detect the evolution of citation cartels over time.
Reference: Liu, J., Xia, F., Feng, X., Ren, J., & Liu, H. (2022). Deep Graph Learning for Anomalous Citation Detection. IEEE Transactions on Neural Networks and Learning Systems.
Summary updated