Citation Graph

Citation graphs are inherently directed (edges point from citing work to cited work) and, in their idealized form, acyclic. A paper can only cite works that already exist, preventing circular reference loops. This temporal constraint creates a natural topological ordering where newer nodes always point to older nodes, enabling chronological traversal of idea evolution. The acyclic property is critical for algorithms like PageRank and HITS that compute node authority scores without getting trapped in cycles.

The importance of a node is quantified using centrality measures:

• Citation Count: The simplest metric, counting incoming edges. Highly susceptible to field-specific norms and self-citation.
• Betweenness Centrality: Measures how often a node lies on the shortest path between other nodes, identifying gatekeeper papers that bridge disparate research communities.
• PageRank: Weights citations by the importance of the citing source, preventing a citation from a low-impact paper carrying equal weight to one from a seminal work.
• Co-Citation Coupling: Two papers are linked if they are both cited by a third paper, revealing thematic similarity.

Citation graphs naturally partition into densely connected clusters representing scientific disciplines, sub-fields, or invisible colleges. Algorithms like Louvain community detection identify these groups by optimizing modularity—maximizing internal edge density while minimizing external connections. These clusters reveal:

• The emergence of new interdisciplinary fields when previously separate clusters begin cross-citing.
• Citation silos where a community primarily cites itself, indicating insularity.
• The lineage of a specific idea as it propagates through a cluster over time.

A citation cascade tracks how a specific idea or finding propagates through the graph over time. Starting from a foundational paper, the cascade maps the tree of subsequent works that cite it directly or indirectly. Analyzing cascade properties—such as depth, branching factor, and temporal velocity—reveals whether a discovery had immediate impact or experienced a sleeping beauty phenomenon, where it lay dormant for years before sudden recognition. This is critical for understanding the true origin of concepts used in AI-generated content.

Not all citations are equal. The edge in a citation graph is binary, but the context surrounding the citation in the citing paper's text carries rich semantic weight. Modern analysis classifies citation intent:

• Background: Acknowledging prior work in the field.
• Methodological: Using a technique or tool from the cited work.
• Supporting: Providing evidence for a claim.
• Contrasting: Disagreeing with or correcting the cited work. Ignoring intent flattens the graph; a critical refutation is treated identically to a supporting reference, distorting influence analysis.

Citation graphs are not static; they evolve with every new publication. Key temporal properties include:

• Preferential Attachment: Highly cited papers are more likely to receive new citations, creating a rich-get-richer effect and a power-law degree distribution.
• Aging and Obsolescence: The probability of a paper being cited decays over time, though the rate varies by field. Foundational methods papers may have exceptionally long half-lives.
• Burst Detection: Identifying sudden spikes in citation activity for a paper or topic, signaling a breakthrough or a trending research area. This temporal signal is vital for maintaining up-to-date source authority scores in generative AI systems.