Citation Graph

In a pure citation graph, edges point from a newer document to an older one, creating a directed acyclic graph (DAG). Time acts as a topological ordering: a document can only cite works that already exist. This property prevents cycles and enables recursive algorithms like PageRank to converge. The acyclic structure also means that the graph's transitive closure reveals the full lineage of influence for any node, tracing ideas back to their origin.

Citation graphs exhibit a bipartite structure between hubs and authorities, as formalized by the HITS (Hyperlink-Induced Topic Search) algorithm.

• Authorities: Documents that are heavily cited by many hubs. They are the definitive sources on a topic.
• Hubs: Documents that cite many high-quality authorities, acting as curated directories. This mutual reinforcement creates a self-correcting ecosystem where genuine authority emerges from the link structure.

A sink node or dangling node is a document that cites no other documents—it has out-degree zero. In algorithms like PageRank, sink nodes act as probability sinks, absorbing rank without redistributing it. This causes rank leak and requires correction via a damping factor (typically d=0.85) that models a random surfer jumping to any node. Real-world citation graphs contain many sinks, such as new papers that haven't been cited yet or terminal reports.

Citation graphs naturally form tightly knit communities through co-citation and bibliographic coupling.

• Co-citation: Two documents are linked if they are both cited by a third document, indicating topical similarity.
• Bibliographic Coupling: Two documents are linked if they cite the same source, revealing shared intellectual foundations. These clustering patterns enable algorithms to identify research fronts, detect emerging fields, and map the evolution of scientific disciplines.

Citation graphs follow a power-law degree distribution, making them scale-free networks. A small number of seminal papers accumulate a disproportionate number of citations (the preferential attachment or 'rich-get-richer' phenomenon), while the vast majority receive few or none. This heavy-tailed distribution means the graph is robust to random node removal but vulnerable to targeted attacks on its hubs. Understanding this skew is critical for normalizing authority scores.

Bridge nodes are documents that cite across disparate communities, connecting otherwise isolated clusters. These nodes have high betweenness centrality and are critical for the diffusion of ideas between fields. In a citation graph, a paper that cites both biology and computer science literature acts as a bridge, enabling cross-pollination. Identifying these nodes helps map the emergence of interdisciplinary fields like bioinformatics.

A computed score reflecting the perceived trustworthiness and expertise of a content source, often derived from a graph analysis of citations and reputation. Unlike simple backlink counts, authority rank evaluates:

• The quality of citing sources, not just quantity
• Recursive weighting where citations from high-authority nodes confer greater rank
• Resistance to manipulation through spam detection and link-farm filtering

This metric is the direct output of algorithms like PageRank applied to the citation graph, serving as a foundational signal for AI models assessing source credibility.

The accuracy with which a generative AI model correctly cites the specific source document or passage that supports a claim in its output. High attribution fidelity means:

• The model points to the exact document that provided the information
• Citations are not hallucinated or fabricated
• The cited passage genuinely supports the generated claim

Poor attribution fidelity manifests as phantom citations—references that look plausible but don't exist—or misattribution, where a real source is cited but doesn't actually contain the claimed information.

An immutable, verifiable record of the sequence of ownership, modifications, and citations for a piece of data, from its origin to its current state. In a citation graph context, provenance chains enable:

• Transitive trust: if A cites B and B cites C, the trust flows through the chain
• Detection of citation laundering, where dubious sources gain credibility through intermediate reputable nodes
• Cryptographic verification that a citation link hasn't been retroactively altered

Provenance chains transform the citation graph from a static snapshot into a temporally-aware trust network.

The ratio of verifiable citations to total claims within a piece of content, serving as a heuristic signal of factual rigor for AI parsers. Key characteristics:

• Calculated as citations ÷ distinct factual assertions
• High density correlates with scholarly or journalistic rigor
• Low density may indicate opinion, marketing content, or unsupported claims

AI models use reference density as a confidence calibration signal—content with sparse citations for bold claims triggers higher epistemic uncertainty, while densely referenced content receives stronger factual grounding scores.

A confidence-boosting indicator derived from multiple independent, authoritative sources corroborating the same factual claim within the citation graph. The strength of a consensus signal depends on:

• Source independence: citations from the same author or organization count less
• Authority diversity: agreement across different high-authority domains
• Graph distance: corroborating sources that are far apart in the citation network carry more weight

When a claim is supported by a dense subgraph of mutually-reinforcing but independent citations, AI models treat it as high-confidence knowledge rather than an isolated assertion.

An NLP task that identifies when two or more statements from different sources provide logically inconsistent information, serving as a negative signal for confidence. In the citation graph, contradiction detection:

• Creates negative edges between conflicting source nodes
• Flags topics where the graph exhibits high disagreement entropy
• Triggers reduced confidence scores for claims in contested regions

Advanced systems use stance detection to classify citation relationships as supporting, refuting, or neutral, enriching the graph beyond simple binary link structures.